Daily Summary for 2026-11-09

daily

Transcript

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

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

Jane: 13 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 Covariance-Calibrated Prediction-Powered Inference.

The day's research: Tom: Fitting complex mixture models is computationally heavy, so developing fast functional clustering is important.

Jane: Creating a method that combines linear mean profiles with first-order structured antedependence covariance gets closed-form conditional covariance updates.

Lu: This method proved significantly quicker than numerical optimization, showing speedups ranging from 91 point five to over 200 times faster in same-R comparisons.

Meng: That speedup means we can perform joint functional clustering much more efficiently without sacrificing model quality or cluster assignments.

Lalam: The framework is built upon combining linear mean profiles on a standardized log niche coordinate with SAD1 covariance.

Tom: This approach showed higher module classification accuracy and lower mean-curve error when compared to methods like K-means or Gaussian mixture models with shared full covariance when the mean structure was correctly specified.

Jane: We also looked at modeling rare events in survival data where extreme value theory is usually applied but complicated by right censoring.

Lu: We proposed a Bayesian survival model that integrates extreme value theory with a hazard-based formulation built upon the extended generalized pareto distribution.

Meng: This allows analysis of moderate and extreme survival times simultaneously while preserving tail behavior.

Lalam: Another piece tackles population surveillance using electronic health records, proposing three Bayesian superpopulation models to estimate type 1 diabetes prevalence in children and adolescents by accounting for unobserved heterogeneity and data sparsity.

Tom: These models showed that the superpopulation-based prevalence estimates were higher than those from the EHR sample alone.

Jane: The analysis of variance approach demonstrated the best estimation performance for that surveillance work.

Lu: Finally, we explored a new minimization method for covariate-adaptive randomization in clinical trials that avoids both shift problems and variance inflation when dealing with unequal allocation and covariate imbalances.

Meng: This new procedure guarantees that marginal imbalances remain bounded in probability while preserving conventional biased-coin allocation probabilities.

Lalam: The most significant development is the proposal for cloudqr which addresses a major practical hurdle in distributed quantile regression where sites cannot easily share raw data.

Tom: This method allows each site to summarize its observations using Gaussian clouds defined by size, center, and covariance such that these clouds perfectly match the score of the raw data.

Jane: A central server then performs the quantile regression on these summarized clouds from all participating sites.

Lu: Simulations and real-world applications show cloudqr significantly outperforms existing distributed methods <a href="/papers/2611.07273">Clouded Representatives and Score Matching for Distributed Quantile Regression</a>.

Tom: This is important because it solves a real-world problem where data privacy or sheer volume prevents direct sharing.

Jane: The work on MSR-CP extends segmented regression to detect gradual shifts in climate data, which are often seen in rainfall patterns rather than abrupt changes.

Lu: The methodology was extended to handle dependent error structures. Simulations on West Bengal monsoon rainfall data suggest it offers improved insights into evolving weather trends <a href="/papers/2611.07858">MSR-CP: A Modified Segmented Regression Approach to Change Point Estimation in Rainfall Patterns in West Bengal</a>.

Tom: This builds upon earlier work that examined existing change point detection methods and their limitations with gradual changes.

Jane: The research on Bayesian detection of competitive-balance eras from match-level data offers a new way to view sports leagues by separating structural ability differences from random match outcomes.

Lu: This framework uses Gaussian auxiliary variables projected onto a graph-fused l1 polytope to infer structural competitive balance <a href="/papers/2611.07335">When Do Sports Leagues Stop Being Competitive? Bayesian Detection of Competitive-Balance Eras from Match-Level Data</a>.

Tom: This is useful because it provides a model-based measure for financial and regulatory debates in sports, showing how team hierarchies can shift over time even when title races are close.

Jane: Fast Functional Clustering via Closed-Form Covariance Updates Functional clustering can be done much faster by using closed-form covariance updates instead of numerical optimization.

Lu: Intensity-Based Estimation of Extreme Value Models for Highly Censored Data via a Latent Poisson Process This Bayesian survival model uses extreme value theory to model censored survival data by unifying the hazard function within a single parametric framework.

Tom: Bayesian Superpopulation Models for Population-Based Diabetes Surveillance Using Multi-Site Electronic Health Record Data These models use Bayesian superpopulation approaches to estimate diabetes prevalence in electronic health records by accounting for unobserved heterogeneity and data sparsity.

Jane: Risk-matched case-control sampling and semiparametric weight calibration for efficient biomarker assessment with binary outcomes This paper proposes a risk-matched case-control design that uses a disease risk score to guide sampling and provides unified inference for association and predictive performance.

Lu: CountScan: Simultaneous Lower Bounds on Historical Watermark Matches over All Text Intervals This method provides finite-sample lower bounds on the number of retained historical watermark matches simultaneously across all text intervals.

Tom: Generalized confidence intervals for the covariate-specific overlap coefficient This work develops a method to construct generalized confidence intervals for the overlap coefficient using linear regression models with normally distributed errors.

Jane: An Improved Minimization Method: Eliminating Shift and Variance Inflation While Maintaining Marginal Balance This paper introduces a new minimization method that avoids the shift problem and variance inflation in covariate-adaptive randomization.

Lu: The Hidden Markov Ising Model with Exogenous Covariates This hybrid model uses an Ising model within a Hidden Markov Model to analyze temporal state changes, such as gun violence activity, by incorporating exogenous variables.

Tom: MSR-CP: A Modified Segmented Regression Approach to Change Point Estimation in Rainfall Patterns in West Bengal This study proposes an extended segmented regression model to detect gradual shifts in rainfall patterns in time series data.

Jane: Clouded Representatives and Score Matching for Distributed Quantile Regression This method proposes a distributed quantile regression technique called cloudqr that allows multiple sites with different data distributions to be combined effectively.

Lu: When Do Sports Leagues Stop Being Competitive? Bayesian Detection of Competitive-Balance Eras from Match-Level Data This Bayesian framework infers structural competitive balance from match outcomes to detect shifts in league structure over time.

Tom: Sample size calculation comparing two survival curves in the presence of nonproportional hazards utilizing the concept of Weighted Relative Time using two different Weibull distributions This paper extends sample size calculations for survival curves with nonproportional hazards by using a concept called Weighted Relative Time.

Jane: Covariance-Calibrated Prediction-Powered Inference Covariance-calibrated prediction-powered inference uses unlabeled prediction scores to calibrate covariance weights in a generalized method of moments estimator.

Tom: The Hidden Markov Ising Model with Exogenous Covariates This hybrid model uses an Ising model within a Hidden Markov Model to analyze temporal state changes, such as gun violence activity, by incorporating exogenous variables.

Jane: Risk-matched case-control sampling and semiparametric weight calibration for efficient biomarker assessment with binary outcomes This paper proposes a risk-matched case-control design that uses a disease risk score to guide sampling and provides unified inference for association and predictive performance.

Lu: CountScan: Simultaneous Lower Bounds on Historical Watermark Matches over All Text Intervals This method provides finite-sample lower bounds on the number of retained historical watermark matches simultaneously across all text intervals.

Tom: Generalized confidence intervals for the covariate-specific overlap coefficient This work develops a method to construct generalized confidence intervals for the overlap coefficient using linear regression models with normally distributed errors.

Jane: An Improved Minimization Method: Eliminating Shift and Variance Inflation While Maintaining Marginal Balance This paper introduces a new minimization method that avoids the shift problem and variance inflation in covariate-adaptive randomization.

Lu: MSR-CP: A Modified Segmented Regression Approach to Change Point Estimation in Rainfall Patterns in West Bengal This study proposes an extended segmented regression model to detect gradual shifts in rainfall patterns in time series data.

Tom: Clouded Representatives and Score Matching for Distributed Quantile Regression This method proposes a distributed quantile regression technique called cloudqr that allows multiple sites with different data distributions to be combined effectively.

Jane: When Do Sports Leagues Stop Being Competitive? Bayesian Detection of Competitive-Balance Eras from Match-Level Data This Bayesian framework infers structural competitive balance from match outcomes to detect shifts in league structure over time.

Lucky paper: 2610.21837: Tom: We’re looking at Covariance-Calibrated Prediction-Powered Inference today, which uses unlabeled prediction scores to calibrate the covariance weight within a Generalized Method of Moments estimator.

Jane: So, it takes those unlabeled prediction scores and uses them to sharpen the covariance weight in a GMM estimator. What exactly does that mean for someone who isn't deep in econometrics?

Lu: It means they are substituting the labeled prediction-covariance contribution with its corresponding unlabeled counterpart within the auxiliary residual metric.

Meng: When we look at deployment, this is interesting because it relies on those prediction scores, which might be generated by a model that’s already trained, so it's an additive step.

Lalam: The paper establishes first-order efficiency under regular fixed-moment conditions and derives exact finite-sample risk identities for linear regression.

Tom: Exact risk identities are strong; what kind of improvement did they show compared to the original GMM-PPI?

Jane: In linear regression simulations with a sample size of five hundred at a design ratio of one to two Covariance-Calibrated Prediction-Powered Inference achieves an MSE that is sixty-eight point nine percent of the supervised MSE.

Lu: That is a substantial reduction when compared to the original GMM-PPI setup.

Meng: Sixty percent reduction is pretty significant for minimizing error at scale, but I wonder how sensitive that sixty-eight point nine percent figure is to the smoothness of our underlying moment function.

Lalam: The construction works with smooth moments and the analysis incorporates exact parameter-risk theory for linear regression alongside empirical evaluation of nonlinear logistic moments.

Tom: So it handles both linear and non-linear moment scenarios, which gives it a broader applicability across different modeling tasks.

Jane: They also have this quadratic decision model that provides a midpoint minimax-regret interpretation specifically for half strength default.

Lu: That interpretation is tied to the exact path analysis, which offers an explicit representation for continuum comparisons and risk approximations regarding half strength default risk.

Meng: If we were deploying this to a real-time system, we'd need to know how much computational overhead that quadratic decision model adds compared to just using the original GMM-PPI without calibration.

Tom: Independent fixed-strength references only yielded minor further gains on the calibration paths studied, which suggests the gain comes primarily from using those unlabeled prediction scores.

Jane: That means relying solely on independent labels isn't as powerful as this calibrated approach is for improving prediction accuracy.

Lu: The paper shows that half strength captures a portion of its potential gain when looking at the exact path analysis for continuum comparisons.

Meng: So, the core mechanism involves using raw targets that preserve the prediction-corrected channel and substituting the labeled prediction-covariance contribution with its unlabeled counterpart.

Lalam: Spectral protection is employed to guarantee positive definiteness, and CC-PPI uses an equal blend of the original covariance and this protected target.

Tom: It’s important that they used spectral protection to ensure the covariance matrix stays positive definite throughout the process.

Jane: This setup allows them to derive those exact finite-sample risk identities under regular fixed-moment conditions.

Lu: They also showed a design-based improvement theorem that accounts for prediction error tolerance, which is a key piece for understanding when this method performs best in practice.

Meng: From an engineering standpoint, if the prediction error tolerance is too wide, does the MSE reduction still hold up under production noise?

Tom: The empirical evaluations on real-data applications demonstrated substantial MSE reductions across all tested tasks at reported sample sizes in both simulations and real-data applications.

Jane: So we see this benefit not just in idealized simulations but also when applied to actual data sets.

Lu: That is the main result showing that Covariance-Calibrated Prediction-Powered Inference consistently produces substantial MSE reductions across all tested tasks at reported sample sizes in both simulations and real-data applications.

Meng: So, if we are running a model on live traffic data, does this calibration process need to be recomputed frequently?

Lalam: The method is constructed to be applicable to smooth moments, which suggests it's well-suited for data where the underlying moment function is continuous rather than just a few discrete points.

Tom: That smoothness requirement limits its use when we are dealing with very noisy or highly discontinuous input data.

Jane: It’s interesting how they balance the exact parameter-risk theory for linear regression with empirical evaluation of nonlinear logistic moments in their analysis.

Lu: That dual approach gives confidence that the theoretical guarantees translate well to more complex, real-world scenarios where things aren't perfectly linear.

Meng: If we move away from linear regression into something non-linear, does the complexity of the auxiliary residual metric become a bigger bottleneck for deployment?

Tom: The paper focuses heavily on linear regression first, so we need to see how those results translate when we try to apply this structure to more complex models.

Jane: Overall, it’s about using unlabeled prediction scores to calibrate covariance weights in a GMM estimator and achieving significant MSE reductions.

Lucky paper: 2611.07335: Tom: Okay, so we are looking at When Do Sports Leagues Stop Being Competitive, this paper is about using a Bayesian framework to infer structural competitive balance from match data.

Jane: The core idea here is projecting Gaussian auxiliary variables onto a graph-fused l one polytope whose radius tells us the total pairwise ability variation across teams. So, how does that projection actually translate into something meaningful for the league structure?

Lu: It’s crucial because it assigns positive probability to exact ties, which lets us form explicit rank tiers for statistically indistinguishable teams and separates genuine ability gaps from just noise in a single season's points table.

Meng: So, if we think about deploying this at scale, the main challenge is defining that graph-fused l one polytope accurately when you have thousands of teams and constant data streams.

Lalam: The framework incorporates an adaptive finite-regime model which permits competitive eras to recur or emerge while team hierarchies remain season specific, which helps capture those recurring shifts we see in sports.

Tom: That idea of the radius equaling the realized dispersion when the polytope constraint binds sounds like a very interpretable measure of league-wide latent ability dispersion.

Jane: It’s more than just looking at points; it’s inferring parity and hierarchy through this probabilistic construction, which is a big shift from retrospective summaries.

Lu: Empirical analysis on twenty-five English Premier League seasons shows a persistent shift toward greater structural inequality starting in the two thousand five/six season, with two thousand ten/eleven being the only later return to a compressed regime.

Meng: That finding is interesting because it suggests that even within the same competitive era, team identities of leading clubs can vary depending on that radius value, which complicates simple structural assumptions.

Tom: So the paper shows that a close title race can coexist with wide league-wide separation and how head-to-head results can separate clubs level on points.

Jane: It moves the assessment from what happened last season to an inference about parity and structural change, which is much more forward looking for forecasting.

Lu: In contrast, the NBA application shows recurrent switching between regimes rather than a single lasting shift, which highlights how different sports have different structural dynamics.

Meng: From an engineering standpoint, the computational diagnostics confirm good cross-chain agreement for most monitored quantities with both datasets favoring a two-state representation, although the Premier League showed greater uncertainty about a third state.

Lalam: The framework suggests that letting the radius depend on covariates like revenue sharing could allow us to estimate associations between regulatory changes and structural balance.

Tom: That’s an interesting implication because it links regulatory shifts directly to how the league's structure is perceived probabilistically.

Lucky paper: 2611.08227: Tom: So we're looking at Fast Functional Clustering via Closed-Form Covariance Updates today, which is a new way to speed up functional clustering by using linear mean profiles and first-order structured antedependence covariance.

Jane: It sounds like they are trying to get around the slow numerical optimization steps that usually plague these kinds of analyses. How does this closed form approach actually work in practice?

Lu: They model the distinct expression patterns using Gaussian mixture components where each feature has a single cluster membership across all conditions, and this whole framework relies on a shared covariance structure.

Meng: That shared covariance structure is key, right? If it's shared across conditions but varies by cluster, does that mean we get better results than models that treat every condition independently?

Lalam: Yes, the model formulation for joint analysis explicitly states that the mean profile for cluster k under condition l is modeled as a linear function of the standardized log niche coordinate with parameters that can vary by both cluster and condition.

Tom: So you have these separate parameters for each cluster and each condition, but they are all tied back to that single shared covariance structure across everything.

Jane: And the covariance itself is modeled using a first-order structured antedependence model called SAD1, which expresses dependence through conditional regressions on earlier ordered measurements.

Lu: That SAD1 structure assumes the covariance blocks are block-diagonal, which captures residual dependence within each condition while keeping things independent across conditions conditional on cluster membership.

Meng: So if a feature is in cluster one, its relationship with another feature in that same cluster is governed by SAD1, but its relationship with a feature in cluster two is independent?

Lalam: Exactly. The estimation procedure uses an expectation-maximization procedure where the M step alternates between weighted least squares mean updates and direct conditional covariance updates.

Tom: That alternating update process sounds like it’s what makes the speedup possible, right? What does that mean for the actual fitting time?

Lu: They claim that this closed-form SAD1 conditional update substantially accelerates covariance estimation by ninety-one point five to over two hundred one point four times in same-R comparisons when solving the conditional objectives.

Jane: That is a massive factor, so if you look at the computational acceleration results for Fast Functional Clustering via Closed-Form Covariance Updates, they show median complete workflow speedups of one hundred fifty-six to two thousand nine hundred thirteen times relative to the original FunClu software across benchmark configurations.

Meng: Two thousand nine hundred thirteen times is huge when you're trying to ship models that need to run quickly at scale. But what about the accuracy side of things?

Lu: When they look at model recovery simulations, they found higher module classification accuracy and lower mean-curve error compared to K-means and a Gaussian mixture with shared full covariance when the mean structure was correctly specified.

Tom: That's good, but you mentioned that performance gain only happens when the mean structure is correctly specified; what happens if our initial guess for the linear function of the log niche coordinate is wrong?

Jane: The paper does point out that curved-mean controls identified the importance of an appropriate mean family, suggesting that getting those means right is a critical part of using this method effectively.

Lalam: The model selection criteria involve jointly selecting the number of clusters K and the degree of covariance parameter sharing S by minimizing the Bayesian information criterion, or BIC.

Lu: They define the total number of free parameters as qK,S equals K minus one plus two KL plus q Sigma S, where q Sigma S depends on that specification S.

Tom: So they aren't just optimizing for speed; they are also optimizing the complexity tradeoff between fitting too many clusters and having a too-simple model.

Jane: The gut and paired leaf–root microbiota datasets were used to illustrate the framework across different numbers of conditions, specifically selecting K equals eleven modules with a BIC of negative two million eight hundred seventeen thousand eight point four zero.

Meng: That number for K is pretty specific; it suggests that for that particular gut multi-omics data set they used, eleven clusters was the best fit according to their metric.

Lalam: And the model selection involves minimizing BIC equals negative two L B K S plus qK S log p.

Tom: It seems like a very thorough approach to finding the right balance between computational efficiency and statistical fit for this functional clustering task.

Lucky paper: 2611.08068: Tom: Alright, we're looking at Eduardo Janotti and his team's work on the Intensity-Based Estimation of Extreme Value Models for Highly Censored Data via a Latent Poisson Process.

Jane: This paper tackles extreme value theory, which is great for modeling rare events, but survival analysis gets complicated when you have right censoring or heterogeneous follow-up.

Lu: They propose a Bayesian survival model that embeds the tail behavior of the extended generalized pareto distribution directly into the hazard function instead of separating threshold exceedances from the bulk of observations.

Meng: That sounds like it’s trying to fix a major numerical instability when you have too much censoring, right?

Tom: Exactly. They use a counting-process framework for this, which is interesting because it connects hazard-based and intensity-based modeling.

Jane: They go through two stages of inference: first approximating the continuous hazard with a piecewise exponential model, and then treating those interval estimates as noisy observations of the underlying generalized pareto distribution.

Lu: They introduce latent evaluation locations to handle the fact that an interval-level hazard estimate doesn't naturally correspond to a single deterministic time point.

Tom: So they have this unobserved location uncertainty built right into the hierarchical model, which is a big move for computational tractability.

Jane: They define the extended generalized pareto distribution using a power transformation applied to the generalized pareto distribution, introducing an additional shape parameter kappa that controls how much it deviates from the standard generalized pareto distribution.

Lu: That parameter kappa allows the extreme value index, ξ, to take negative, zero, or positive values corresponding to bounded-tail, exponential-tail, and heavy-tailed behavior without needing a priori restrictions.

Tom: If you have that flexibility for the tail index xi, it means the model can adapt itself based on how extreme your data actually gets.

Jane: They address discretization bias from using piecewise exponential approximations by considering two mitigation strategies: selecting a deterministic location where the hazard matches the average hazard over an interval, or using an adaptive partition refinement.

Lu: The approach they actually adopted is a latent location formulation, treating that representative location as an unobserved quantity with uncertainty incorporated directly into the hierarchical model structure.

Meng: What about those simulation results? Does it hold up when the data generating mechanism is outside the assumed generalized pareto family?

Tom: They ran two simulation studies. The first looked at how many bins you use on a correctly specified generalized pareto distribution, showing that a small number of bins, like M equals five or M equals ten gives reliable inferential performance.

Jane: The second study checked robustness against model misspecification by generating survival times from a lognormal distribution and showed the proposed model could still provide a close approximation to that underlying distribution.

Lu: When applied to the heart failure dataset, they looked at long-term survival under substantial right censoring, and they found evidence of a difference in the scale parameter between male and female patients.

Tom: That’s interesting because their posterior credible interval for the tail effect included zero, which suggests limited evidence of a difference in asymptotic tail behavior between the two groups.

Jane: The conclusion is that this methodology gives a flexible and computationally tractable alternative for estimating generalized pareto distributions under right censoring by using an intensity-based piecewise poisson representation.

Lu: They also identified future research directions focused on systematically investigating the adaptive partitioning procedure and data-driven selection of the number of intervals.

Lucky paper: 2611.08238: Tom: So we're looking at Risk-matched case-control sampling and semiparametric weight calibration for efficient biomarker assessment with binary outcomes. It’s a pretty heavy methodology, focusing on two-phase designs to reduce the cost of collecting those expensive measurements.

Jane: The core idea here is using a scalar disease risk score from Phase I variables to define matching strata that guide the sampling in Phase II, which aims to give you efficient inferences for both association parameters and predictive performance metrics like positive predictive value and the area under the receiver operating characteristic curve.

Lu: The paper outlines a three-step process, starting with defining the logistic regression model for binary outcome Y using those coefficients beta zero, beta one, and beta two.

Meng: And then they focus on developing inverse probability weighted estimators for those betas, which accounts for unequal inclusion probabilities by assigning each sampled subject a weight equal to one minus the inclusion probability.

Tom: But they go further than just basic IPW; they propose a two-step procedure that combines model-assisted weight calibration with stratum-specific weight scaling to stabilize those weights and improve efficiency.

Jane: That calibration step is key because it adjusts the original IPW weights using Phase I information, which helps address the issue of unequal inclusion probabilities in a more targeted way.

Lu: The simulation results show that this proposed Risk-matched case-control design generally achieves the highest or near-highest efficiency across various sampling designs, especially for estimating those regression coefficients.

Tom: And they specifically point out that parameter-specific calibration consistently improved efficiency over basic IPW within each design, with one metric showing a substantial improvement in AUC efficiency recovering approximately eighty-five to ninety percent of the Phase I estimator's efficiency compared to only twenty to fifty percent for the standard IPW.

Jane: That recovery factor is quite compelling, especially when you look at how they compare it against the corresponding Phase I estimator, which is what we're trying to beat with these two-phase designs.

Meng: When applying this methodology to their application on oral precancer progression risk prediction using data from a Taiwan cohort study, they compared three analysis settings: RMCC with IPW, RMCC with the proposed weight calibration, and CC with IPW.

Lu: The results from that application showed that combining the Risk-matched case-control design with those calibrated weights provides substantial efficiency gains over the standard CC design for both estimating association parameters and AUC.

Tom: So it seems like tailoring that calibration procedure to match the specific parameter or prediction measure you're interested in is what drives those efficiency gains, which is a big practical point for real studies.

Jane: That leads into their discussion, where they conclude that those gains depend on how well the calibration variables align with the target parameters you are trying to estimate.

Meng: They also suggest future work could look at extending these Risk-matched case-control designs to time-to-event outcomes by using Cox regression risk models.

Lu: The authors acknowledge Drs. Anil Chaturvedi and Kelly Yu for providing the data from that oral precancer natural history study in Taiwan, which is crucial context for their application.

Tom: It’s interesting how they are tying the efficiency gains directly back to the alignment between their calibration variables and what they actually want to predict or estimate, rather than just applying a blanket correction.

Jane: And on a practical note, that means researchers need to think carefully about what parameter they prioritize when setting up their Phase I design if they want those specific efficiency boosts.

Lu: That's the main direction for future research they point toward, focusing on how this RMCC framework can be adapted for different types of survival data structures.

Meng: From a deployment standpoint, the focus on achieving high efficiency across multiple parameters like association and AUC is what makes this method attractive for cost-effective screening tests where you need both predictive power and good statistical grounding.

Tom: It’s a lot of machinery in there, but the number showing that CIPW substantially improves AUC efficiency over IPW by recovering eighty-five to ninety percent of the Phase I estimator’s performance is a solid piece of evidence.

Jane: Exactly, it shows that even with complex two-phase designs and unequal probabilities, you can still get much better results than just using standard inverse probability weighting.

More episodes

← Home