Overstuffed sandwiches and separation anxiety: finite-sample variance estimation for penalized GEE with near-separated binary data
Listen
Radio episode about this paper
Transcript
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: "Overstuffed sandwiches and separation anxiety".
Paper summary: Jane: So, wrapping up the discussion on this paper from "Overstuffed sandwiches and separation anxiety: finite-sample variance estimation for penalized GEE with near-separated binary data." It sounds like the main point is that when you have these challenging longitudinal binary data sets that are close to being separated, standard ways of estimating variance can be too aggressive or biased in small samples >
Tom: Exactly. The paper doesn't just point out the problem; they give us a specific solution called V-hat AR, which combines the score-level leverage correction with a controlled finite sample upward translation >
Lu: They establish first-order asymptotic theory for penalized GEE along convergent interior-root sequences and also provide a matrix characterization of that parameter-specific overcorrection induced by full leverage adjustment >
Meng: The implication for us in AI deployment is that when we see these small, rare event scenarios, we should be looking at how the variance correction is being applied because it directly impacts the reliability of our inference >
Lalam: For treatment effect inference in those low event, small N settings they studied here, the authors recommend V-hat AR with an N minus p Wald t test as a conservative alternative when standard leverage corrections remain anti-conservative >
Jane: It's about providing a concrete piece of advice for researchers working in these areas, showing that there is a better way to handle the uncertainty when you can't rely on the large sample approximations >
Conclusion: Tom: So, we're wrapping up this look at the paper "Overstuffed sandwiches and separation anxiety: finite-sample variance estimation for penalized GEE with near-separated binary data."
Jane: The title sounds... a bit silly, doesn't it? Overstuffed sandwiches and separation anxiety.
Lu: It’s a metaphor for how when your data is almost completely separated, the math gets really weird.
Meng: What does that actually mean for us when we're trying to run these models in production?
Lalam: Basically, they found a way to fix how they estimate the uncertainty—the variance—when you've got this kind of binary data that’s very close to being perfectly split.
Tom: So, instead of just using standard methods that might be totally off in small samples, they built a new estimator called V-hat AR.
Jane: They use something called V-hat AR to make sure the estimates are more reliable when the sample size is small and the events are rare.
Lu: The authors show some solid math here, establishing how this works along specific sequences of roots in their equations, which gives us a strong foundation for it.
Meng: I'm interested in what they say about that overcorrection they find with full leverage adjustment—is that something we need to worry about when we scale up?
Lalam: Yeah, Theorem two characterizes that parameter-specific overcorrection, showing exactly how much the estimation gets pushed too far when you adjust for all the leverage.
Tom: And in practice, for those low event, small sample settings they tested—like on a trial with ten events—V-hat AR gave them results that were conservative or almost exactly what they expected.
Jane: That’s huge because it means we have a way to trust our treatment effect inference even when the data looks messy and sparse.
Lu: They recommend using V-hat AR with an N minus p Wald t test for those specific scenarios, which is a solid recommendation based on their finite sample performance.
Meng: So, it's not some perfect solution yet, but it gives us a much safer tool to deploy when we can't afford to be wrong in the small N cases.
Lalam: Exactly. It’s a conservative alternative when standard leverage corrections fail you in these specific conditions.
Tom: The paper boils down to giving researchers a concrete way to handle variance estimation for penalized GEE under near-separated binary data, focusing on what happens when the sample size is tiny and the events are rare.
Jane: So we’ve seen how they built this tool and what it does for small studies. Where does this leave us when we think about bigger AI problems?
University of Washington · Institute of Statistical Research and Training, University of Dhaka
stat.ME
Submitted: 2026-04-20
Updated: 2026-11-16
Comments: 69 pages, 8 figures, 21 tables. Includes supplementary appendices in the same PDF
Code: https://github.com/awanafiaz/pgeeVar
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 92/100
The gist: Penalized generalized estimating equations (PGEE) stabilize point estimation for long-term binary data under near-separation, but inference still depends on how the sandwich variance is corrected.
Key concepts
- Penalized Generalized Estimating Equations (PGEE)
- PGEE is a statistical method used for binary data that incorporates penalties to stabilize point estimates. It is particularly useful when the data exhibits near-separation, meaning the groups are somewhat distinct, which helps in estimating long-term effects.
- VˆAR
- VˆAR is a proposed variance estimator that improves upon standard methods. It combines score-level leverage correction with a finite-sample upward translation term. This combination aims to provide more accurate and conservative estimates of the variance, especially when standard corrections are not fully effective.
- Leverage Correction
- Leverage correction is a technique used in variance estimation to account for the influence of extreme observations (high leverage) on the estimated variance. In this paper, it is applied at the score level to adjust for bias introduced by these influential data points in PGEE models.
- Near-Separation
- Near-separation describes a situation in binary data where the two groups being compared are relatively distinct or separated. This condition is important because it allows penalized GEE methods to stabilize the point estimation of long-term effects, making them applicable even when groups are not perfectly separated.
Terminology
Summary
Penalized generalized estimating equations (PGEE) stabilize point estimation for long-term binary data under near-separation, but inference still depends on how the sandwich variance is corrected. p. 0 Overstuffed sandwiches and separation anxiety:
finite-sample variance estimation for penalized GEE
with near-separated binary data
Awan Afiaz1,2,* and M. Shafiqur Rahman2,†
1Department of Biostatistics, University of Washington
2
Institute of Statistical Research and Training, University of Dhaka
Abstract
Penalized generalized estimating equations (PGEE) stabilize point estimation for
longitudinal binary data under near-separation, but inference still depends on how
the sandwich variance is corrected. Existing corrections for PGEE can overadjust in
high-leverage directions, require restrictive pooling assumptions, or add global regularization without explaining the bias. We establish first-order asymptotics for PGEE
along convergent interior-root sequences and derive a matrix characterization of the
parameter-specific overcorrection induced by full leverage adjustment.Where the paper says it
Where the paper says it
Methodology for Bias Correction
The paper introduces VˆAR, a proposed variance estimator that combines score-level leverage correction with controlled finite-sample inflation <ref:2604.18863#pg26>. This estimator keeps the score-level leverage correction and adds a finite-sample upward translation dominated at first order by the finite-sample multiplier, with a smaller centering term <ref:2604. p. 0 Finite-sample
calibration is limited by both mean bias and the variability of leverage-corrected variance estimates. We propose a variance estimator that keeps the score-level leverage
correction and adds a finite-sample upward translation dominated at first order by the
finite-sample multiplier, with a smaller centering term. We propose
Vˆ
AR, which combines score-level leverage correction with controlled finite-sample inflation,
and derive bias expressions and consistency under estimator-specific conditions (Propositions 1 and 2). We compare fourteen estimators across 148 simulation scenarios and two
applied datasets.Where the paper says it
Where the paper says it
Asymptotic Theory for PGEE
The authors establish first-order asymptotic theory for PGEE along convergent interior-root sequences (Lemma 1 and Theorem 1) <ref:2604.18863#pg7>. They characterize parameter-specific leverage overcorrection under correctly specified working covariance (Theorem 2) <ref:2604. p. 3 We establish first-order asymptotic theory for PGEE along convergent interior-root sequences (Lemma 1 and Theorem 1) and characterize parameter-specific leverage overcorrection under correctly specified working covariance (Theorem 2). The balanced intercepttreatment case identifies the role of the smaller treatment arm (Corollary 1). It does not resolve the finite-sample paths on
which PGEE fails to converge. It shows only that when an interior root is selected, the
penalty does not alter the first-order limit. We propose
Vˆ
AR, which combines score-level leverage correction with controlled finite-sample inflation,
and derive bias expressions and consistency under estimator-specific conditions (Propositions 1 and 2). We compare fourteen estimators across 148 simulation scenarios and two
applied datasets.Where the paper says it
Where the paper says it
Where the paper says it
Finite-Sample Performance and Calibration
The authors propose VˆAR as a conservative alternative when standard leverage corrections remain anti-conservative <ref:2604.18863#pg8>. In the low-event, small-N non-pooling PGEE settings studied here, VˆAR yields conservative or near-nominal type I error for treatment-effect inference at N = 10, where existing leverage corrections remain anti-conservative and pooling estimators are inapplicable to unbalanced designs <ref:2604. p. 3 We compare fourteen estimators across 148 simulation scenarios and two
applied datasets. In the low-event, small-N non-pooling PGEE settings studied here, Vˆ
AR
yields conservative or near-nominal type I error for treatment-effect inference at N = 10,
where existing leverage corrections remain anti-conservative and pooling estimators are inapplicable to unbalanced designs. The middle matrix for VˆMD is P
i fif
T
i with fi = DT
i V
−1
i
(I − Hii)
−1
ri
. Vˆ
AR differs from
VˆMD only through the multiplicative factor cN = FPC · Bessel and the mean-centering
by ¯f.Where the paper says it
Where the paper says it
Application Results
The authors apply their methods to two applications, the toenail onychomycosis trial and contagious bovine pleuropneumonia data <ref:2604.18863#pg5>. In the toenail trial, VˆAR gives p = 0. p. 0 In simulations, the proposed estimator gives conservative or near-nominal type I error in settings with low event rates
and few subjects, including designs with ten subjects, where several standard corrections remain anti-conservative and pooling estimators are unavailable for unbalanced
designs. Applications to toenail trial and contagious bovine pleuropneumonia data
illustrate how the choice of variance estimator affects standard errors and inference in
balanced and unbalanced small-sample designs. We compare fourteen estimators across 148 simulation scenarios and two
applied datasets. In the low-event, small-N non-pooling PGEE settings studied here, Vˆ
AR
yields conservative or near-nominal type I error for treatment-effect inference at N = 10,
where existing leverage corrections remain anti-conservative and pooling estimators are inapplicable to unbalanced designs. In the GEE setting, I0 coincides with the true information only when the working correlation is correctly specified. Under misspecification, the penalty preserves the O(1) bound
of Lemma 1. In the GEE setting, I0 coincides with the true information only when the working correlation is correctly specified. Under misspecification, the penalty preserves the O(1) bound
of Lemma 1. We compare fourteen estimators across 148 simulation scenarios and two
applied datasets. In the low-event, small-N non-pooling PGEE settings studied here, Vˆ
AR
yields conservative or near-nominal type I error for treatment-effect inference at N = 10,
where existing leverage corrections remain anti-conservative and pooling estimators are inapplicable to unbalanced designs.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Conclusion on Variance Estimators
The authors conclude that VˆAR and VˆMD share the score-level (I − Hii)−1 correction, which eliminates the selfshrinkage bias of fitted residuals <ref:2604.18863#pg5>. For treatment-effect inference in the low-event, small-N non-pooling PGEE settings studied here, we recommend VˆAR as a conservative alternative when standard leverage corrections remain anti-conservative <ref:2604. p. 3 We compare fourteen estimators across 148 simulation scenarios and two
applied datasets. In the low-event, small-N non-pooling PGEE settings studied here, Vˆ
AR
yields conservative or near-nominal type I error for treatment-effect inference at N = 10,
where existing leverage corrections remain anti-conservative and pooling estimators are inapplicable to unbalanced designs. bd
†Email: shafiq@isrt. bd
1
arXiv:2604.Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Implement a variance estimator called V AR that combines score-level leverage correction with controlled finite-sample inflation to provide conservative treatment effect inference in low-event, small-N non-pooling settings. This system can identify when standard leverage corrections remain anti-conservative and suggest alternative inference procedures such as permutation tests when all non-pooling estimators produce inflated type I error at very low event rates.
-
Develop a framework for characterizing parameter specific overcorrection induced by full leverage adjustment using the matrix characterization Theorem 2. This allows the system to diagnose how different variance estimators, particularly those applying full (I minus Hii) inverse correction, overcorrect in high-leverage directions near separation regimes.
-
Create a tool to compare the performance of various GEE variance estimators across 148 simulation scenarios and two applied datasets by evaluating them at the penalized GEE fit. This system can rank estimators based on their type I error control and median standard errors relative to simulation standard errors, specifically identifying when V AR is superior in terms of conservative inference.
-
Integrate a mechanism to assess working correlation misspecification by comparing results across different true correlation structures (exchangeable vs AR(1)) using Figure 2. This allows the system to confirm that the conclusions drawn from correctly specified working correlations are not artifacts of specifying exchangeable or AR(1) models when they are actually correct.
-
Implement a diagnostic to determine coefficient level leverage sensitivity by calculating the ratio rho s = [Delta Blev Delta]ss / Delta ss. This allows the system to quantify how much the variance estimate for a specific coefficient is influenced by leverage, aiding in understanding coefficient-specific sensitivity under correctly specified working covariance and frozen first-order residual expansion.
Abstract
Penalized generalized estimating equations (PGEE) stabilize point estimation for longitudinal binary data under near-separation, but inference still depends on how the sandwich variance is corrected. Applications to toenail trial and contagious bovine pleuropneumonia data show that the existing estimators can disagree on significance in small samples, and that several cannot be computed when subjects have unequal numbers of observations. We study which variance estimator to use for PGEE in small longitudinal studies with rare events. We establish first-order asymptotics for PGEE along convergent interior-root sequences and derive a matrix characterization of the parameter-specific overcorrection induced by full leverage adjustment. We compare thirteen existing estimators with a proposed estimator that combines full inverse leverage correction, score centering and finite-sample scaling. With ten subjects, only the proposed estimator and Morel's control the type I error rate without assuming that all subjects share a correlation structure, and the proposed estimator is closer to the nominal level. In small samples, its scaling removes the excess rejections left by leverage correction and its centering offsets part of the resulting conservatism. In larger samples the leverage-corrected estimators behave alike. Estimators that average residual covariance across subjects also control the type I error rate, in small and larger samples, but they require a shared correlation structure and equal numbers of observations per subject, which may not always be feasible.
Related papers
- Learning Robust Treatment Rules for Censored Data
- Neural Estimation of Summary Statistics for Simulation-Based Inference with Non-Asymptotic Guarantees
- Spectral divide-and-conquer MCMC for long stationary time series
- Improving the Efficiency of Subgroup Analysis in Randomized Controlled Trials with TMLE
- Personalized Federated Vector Autoregression with Personalization Diversity
- Doubly Robust and Efficient Calibration of Prediction Sets for Right-Censored Time-to-Event Outcomes