Nesting or resolution? Facility proxies for variance estimation versus confounding adjustment in claims data without facility identifiers

arXiv:2610.24679 · stat.ME, stat.AP · Submitted 2026-10-17 · Read on arXiv

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: We're looking at the paper: "Nesting or resolution? Facility proxies for variance estimation versus confounding adjustment in claims data without facility identifiers".

Jane: Here's the gist of the paper. No single facility proxy serves every purpose: nesting governs variance estimation, and resolution governs adjustment for accreditation-linked facility characteristics and assessment of treatment availability Facility identifiers are often removed during de-identification in claims databases, yet they are needed for variance estimation, confounding adjustment, and assessing treatment availability The paper examines whether accreditation-gated billing items can build strata standing for hospitals and how to construct a facility proxy for each use The study analyzed 1,730,736 adults undergoing surgery under general anaesthesia in the Medical Data Vision database, Japan (February 2019–July 2023) Marker-pattern strata were built from 66 accreditation-gated receipt codes (markers), assuming markers are observed with imperfect sensitivity and near-perfect specificity The framework relies on four conditions: some billable items can be billed only by hospitals accredited for them, the markers remain after de-identification, observation of the markers errs in one direction, and hospitals are separable Candidate groupings included observed patterns (marker-pattern strata), latent class analysis of binary marker indicators, and cross-classification of every supplied facility field (supplied-attribute cells) Performance in Variance Estimation The variance explained by the marker-pattern strata decreased with the extent to which a variable was patient- rather than hospital-determined, from 0.921 for accreditation for generic drug use to 0.036 for patient sex For variance estimation, supplied-attribute cells kept the type I error rate at 4.8% with a wild cluster bootstrap against 5.4% for the true identifier and 39% when clustering was ignored Nesting governs variance estimation Performance in Adjustment and Availability The marker pattern removed 51% of the confounding related to accreditation-linked facility characteristics against 8% for the cells when evaluated by pairing a real facility-level confounder with a simulated null exposure For treatment availability, the proportion of patients in groups mixing hospitals with and without relevant accreditation was 49% for the marker pattern against 98% for the cells Resolution governs adjustment and treatment availability Where an analysis requires adjustment for, or clustering on, the treating hospital but no identifier is supplied, the supplied-attribute cells can carry the standard errors, and marker-pattern strata are worth considering for facility characteristics aligned with the accreditation system The framework's applicability is governed by specificity, sensitivity degradation, and separability The study concludes that nesting governs variance estimation and resolution governs adjustment for facility characteristics Limitations A limitation is that sensitivity in the application database could not be measured directly, and results for adjustment and treatment availability rest on simulated null exposures paired with real confounders Furthermore, strata of median size 10 may often be uninformative for binary exposures and rare outcomes A hospital that never bills leaves no marker and does not lower the measured sensitivity, potentially causing issues in adjustment and treatment availability The study also notes that a proxy that merges hospitals inherits properties, whereas one that splits them reproduces the inflation caused by misspecified cluster membership Data and Code Availability The two claims databases were obtained under commercial licences from Medical Data Vision Co., Ltd.

Tom: and DeSC Healthcare, Inc.

Paper summary: Tom: So, wrapping up this paper on "Nesting or resolution? Facility proxies for variance estimation versus confounding adjustment in claims data without facility identifiers," the authors are essentially saying that there's no one-size-fits-all solution for using leftover facility information in claims data.

Jane: They're arguing that you should choose your tool based on what you need to do: if your primary goal is estimating variance, nesting is the better choice because it handles the structure of hospital outcomes well.

Lu: Conversely, if you are trying to adjust for things like accreditation status or treatment availability, resolution is what you need because it handles that type of confounding more effectively.

Meng: The real world implication here is that when we're building models at scale, we can't just assume one method works for everything; the choice between nesting and resolution depends entirely on the specific question you are asking about those facility proxies.

Tom: They show that using marker-pattern strata is a worthwhile consideration when you are looking at facility characteristics that align with a specific accreditation system, even without having direct hospital identifiers in your dataset.

Jane: It’s important to remember that this framework has its limits, though; the authors admit that sensitivity degradation and separability are things they have to keep in mind when applying it to real data.

Lalam: The paper provides a concrete way to build these proxies from billing records under the assumption that some billable items are billed only by hospitals accredited for them, and those markers remain after de-identification.

Tom: So, if you're working on healthcare data where you have lost identifiers but need to account for hospital effects or accreditation differences, this paper gives you a roadmap for picking the right proxy tool.

Conclusion: Tom: So, this paper is about finding the right way to use hospital information when you don't have those direct hospital IDs in your data for things like insurance claims.

Jane: Right, so the title asks whether nesting or resolution is better when you're trying to figure out variance versus adjusting for other factors.

Lu: Basically, they look at how you can build a proxy—a stand-in for the facility—from accreditation codes and see what that proxy does.

Meng: I wonder if this is actually deployable? Because building those strata sounds like it would be a huge lift to set up on my end.

Lalam: The study uses one point seven million adults from surgery data in Japan and creates these marker-pattern strata based on sixty-six accreditation codes.

Tom: And the results show that nesting is good for estimating variance, while resolution works better for adjusting those confounding factors related to the hospital itself.

Jane: That means if you just want a better picture of how much things vary between patients, nesting is your bet.

Lu: But if you're trying to adjust for whether a patient had treatment available or other hospital-specific traits, resolution handles that adjustment more effectively.

Meng: So it’s like having two different tools in your toolbox depending on the job you need done with the data.

Lalam: The study found that nesting removes fifty-one percent of the noise from accreditation characteristics when you're looking at variance explained, compared to only eight percent for those cells.

Tom: And for treatment availability, resolution reduced the mixing of hospitals with and without accreditation from nearly ninety-nine percent down to about forty-nine percent.

Jane: It points toward a framework where the choice depends entirely on whether you need a better variance estimate or a better adjustment for hospital effects.

Lu: The main thing is that this approach shows how to use facility data strategically even when the original identifiers are gone, provided your assumptions about sensitivity hold up.

Graduate School of Informatics, Osaka Metropolitan University · German Research Center for Artificial Intelligence (DFKI) · Department of Public Health, Graduate School of Medicine, Osaka Metropolitan University · Department of Medical Statistics, Graduate School of Medicine, Osaka Metropolitan University · Department of Computer Science, RPTU Kaiserslautern-Landau

stat.ME, stat.AP

Submitted: 2026-10-17

Updated: 2026-11-04

Comments: 11 pages, 4 figures, 3 tables, 2 ancillary files (Additional files 1 and 2). v2

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 81/100

The gist: No single facility proxy serves every purpose: nesting governs variance estimation, and resolution governs adjustment for accreditation-linked facility characteristics and assessment of treatment…

Key concepts

Marker-pattern strata
These are groupings created by combining observed patterns of billing codes (markers) from accreditation-gated receipt codes. The study tests if these patterns can serve as a proxy for the actual hospital, assuming the markers are observed imperfectly.
Nesting
This concept describes how grouping data affects variance estimation. The paper found that nesting is effective for variance estimation when analyzing hospital-level characteristics, showing that this method is appropriate depending on whether you are trying to estimate patient- or hospital-driven variation.
Resolution
This concept relates to adjusting for confounding factors and assessing treatment availability. The study concluded that resolution governs adjustment because it allows researchers to effectively remove the influence of accreditation-linked facility characteristics from the data.

Terminology

Summary

No single facility proxy serves every purpose: nesting governs variance estimation, and resolution governs adjustment for accreditation-linked facility characteristics and assessment of treatment availability

Background and Motivation

Facility identifiers are often removed during de-identification in claims databases, yet they are needed for variance estimation, confounding adjustment, and assessing treatment availability The paper examines whether accreditation-gated billing items can build strata standing for hospitals and how to construct a facility proxy for each use

Methodology

The study analyzed 1,730,736 adults undergoing surgery under general anaesthesia in the Medical Data Vision database, Japan (February 2019–July 2023) Marker-pattern strata were built from 66 accreditation-gated receipt codes (markers), assuming markers are observed with imperfect sensitivity and near-perfect specificity The framework relies on four conditions: some billable items can be billed only by hospitals accredited for them, the markers remain after de-identification, observation of the markers errs in one direction, and hospitals are separable Candidate groupings included observed patterns (marker-pattern strata), latent class analysis of binary marker indicators, and cross-classification of every supplied facility field (supplied-attribute cells)

Performance in Variance Estimation

The variance explained by the marker-pattern strata decreased with the extent to which a variable was patient- rather than hospital-determined, from 0.921 for accreditation for generic drug use to 0. p. 0036 for patient sex For variance estimation, supplied-attribute cells kept the type I error rate at 4. p. 08% with a wild cluster bootstrap against 5. p. 04% for the true identifier and 39% when clustering was ignored Nesting governs variance estimation

Performance in Adjustment and Availability

The marker pattern removed 51% of the confounding related to accreditation-linked facility characteristics against 8% for the cells when evaluated by pairing a real facility-level confounder with a simulated null exposure For treatment availability, the proportion of patients in groups mixing hospitals with and without relevant accreditation was 49% for the marker pattern against 98% for the cells Resolution governs adjustment and treatment availability

Conclusion on Proxy Utility

Where an analysis requires adjustment for, or clustering on, the treating hospital but no identifier is supplied, the supplied-attribute cells can carry the standard errors, and marker-pattern strata are worth considering for facility characteristics aligned with the accreditation system The framework's applicability is governed by specificity, sensitivity degradation, and separability The study concludes that nesting governs variance estimation and resolution governs adjustment for facility characteristics

Limitations

A limitation is that sensitivity in the application database could not be measured directly, and results for adjustment and treatment availability rest on simulated null exposures paired with real confounders Furthermore, strata of median size 10 may often be uninformative for binary exposures and rare outcomes A hospital that never bills leaves no marker and does not lower the measured sensitivity, potentially causing issues in adjustment and treatment availability The study also notes that a proxy that merges hospitals inherits properties, whereas one that splits them reproduces the inflation caused by misspecified cluster membership

Data and Code Availability

The two claims databases were obtained under commercial licences from Medical Data Vision Co., Ltd. and DeSC Healthcare, Inc. p. 1

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

  1. Use facility proxies for variance estimation by employing supplied-attribute cells when analyzing outcomes correlated within hospitals, as nesting governs variance estimation. This allows AI models to maintain cluster-robust standard errors when a true facility identifier is unavailable.

  2. Implement resolution strategies for adjustment and treatment availability by using marker pattern strata, as resolution governs adjustment for accreditation-linked facility characteristics and assessment of treatment availability. This enables AI systems to accurately assess the impact of specific hospital characteristics on outcomes without requiring a perfect hospital identifier.

  3. Develop a flexible framework that dynamically selects the grouping method based on the analytical goal: use supplied-attribute cells for variance estimation, marker pattern strata for adjustment, and marker pattern strata for treatment availability. This allows AI systems to choose the optimal proxy strategy depending on whether they need to estimate uncertainty or adjust for confounding/availability.

  4. Incorporate sensitivity and specificity assessments into model validation using simulations where sensitivity degraded inference gradually, allowing AI systems to quantify how imperfect marker observation affects their performance estimates. This helps in selecting appropriate proxies based on the expected quality of data identifiers.

Related papers