📊 statistics

Estimating Association Between Paired Outcomes in Clustered Data with Informative Subgroup Size

This paper proposes weighted estimators and associated inference methods to accurately estimate marginal associations between paired tooth-level outcomes in clustered dental data, specifically addressing biases caused by informative cluster sizes and subgroup structures.

Original authors: Owen Visser, Somnath Datta

Published 2026-05-18

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Owen Visser, Somnath Datta

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Counting Teeth in a Crowded Room

Imagine you are trying to figure out how two things are related in a group of people: how many teeth they have lost (Periodontal disease) and how many cavities they have (Caries).

In a normal study, you might just count every single tooth from every person and run the numbers. But here is the problem: People don't all have the same number of teeth. Some people have a full set of 28 teeth; others might only have 10 because they lost the rest due to past illness.

If you just count every tooth equally, the people with 28 teeth get to "vote" 28 times, while the people with 10 teeth only get to vote 10 times. This creates a bias. It's like a room where the loudest people (those with more teeth) drown out the quieter ones. The paper argues that in dentistry, the number of teeth a person has isn't random; it's often a sign of how sick they were in the past. So, simply counting every tooth gives a distorted picture of the relationship between gum disease and cavities.

The Core Problem: "Informative" Groups

The authors call this Informative Cluster Size (ICS).

The Cluster: The patient (the person).
The Units: The teeth inside that person's mouth.
The Issue: The size of the cluster (how many teeth) tells you something about the outcome (how sick the person is).

If you ignore this, your math is like trying to measure the average height of a basketball team by counting every player, but accidentally counting the tall players three times and the short players once.

The New Solution: Three New Ways to "Vote"

The paper proposes three new ways to weight the data so that every person gets an equal say, regardless of how many teeth they have left. They imagine a hypothetical game where we pick teeth to measure, but we change the rules of the game to fix the bias.

Think of the teeth in a mouth as different flavors of ice cream (e.g., Chocolate, Vanilla, Strawberry). Some people have a huge bowl with 20 scoops of Chocolate and 0 of Vanilla. Others have 1 scoop of each.

The authors suggest three different ways to pick a "representative" scoop from each person's bowl so that the Chocolate-heavy bowls don't dominate the results:

Population Pair Weights (PPW): Imagine you have a menu of every possible flavor combination that exists in the whole world. You pick a flavor combination at random (say, "Chocolate & Vanilla"), and then you look for a person who has that specific combo. If a person has 10 scoops of Chocolate, they are just as likely to be picked as someone with 1 scoop, because the "menu" dictates the choice, not the bowl size.
- The Catch: This method is very sensitive. If the "menu" is too specific, the math gets shaky and can produce wrong answers (high bias).
Observed Pair Weights (OPW): Imagine you only look at the flavors actually present in that specific person's bowl. If they have 20 scoops of Chocolate, you pick one of those 20. But, you make sure that the "Chocolate" category only gets one vote total for that person, no matter how many scoops they have.
- The Catch: This is safer than PPW, but still not perfect in every scenario.
Marginally Observed Pair Weights (MOPW): This is a middle-ground approach. You look at the flavors present in the bowl and create a "grid" of all possible combinations of those flavors, even if some combinations don't actually exist in that bowl. You then pick from this grid.
- The Catch: Like the others, it changes the answer depending on how you define the "flavors."

The Test: Did the New Math Work?

The authors ran a massive computer simulation (like a video game) to see if these new rules actually fixed the problem. They created fake patients with fake teeth and fake diseases, knowing the "true" answer beforehand.

The Results:

The Good News: They created a new test (called the ISS Test) that acts like a "smoke detector." It can successfully tell you when the number of teeth is biased. It worked very well in the simulation.
The Bad News: The new weighting methods (PPW, OPW, MOPW) did not consistently give the correct answer. In fact, sometimes they made the answer worse or less reliable than the old, simple method.
The Winner: The old method, called Cluster Weighting (CW), which simply gives every person one vote regardless of their tooth count, turned out to be the most stable and reliable in the simulations.

The Real-World Test: NHANES Data

The authors applied their methods to real data from the NHANES (a massive US health survey). They looked at the link between gum disease and cavities in real people's mouths.

What they found: There is a small, positive link between gum disease and cavities (if you have one, you likely have the other).
The Twist: The strength of this link changed depending on which "weighting" method they used.
- When they looked at filled teeth (teeth that had been treated), the new methods showed a big difference compared to the old method. This suggests that for treated teeth, the number of teeth a person has is very informative.
- For decayed teeth (untreated cavities), the methods were more similar.

The Bottom Line

The paper concludes with a very cautious message:

We have a better "smoke detector": We now have a good way to test if the number of teeth is biasing our results.
We don't have a magic bullet yet: The fancy new ways of weighting the data (PPW, OPW, MOPW) are interesting ideas, but they aren't perfect. They can change the results significantly, but we don't know if they are changing them for the better or just making them different.
Stick to the basics for now: Until the new methods are proven more reliable, the safest bet is to treat every patient as one equal vote (the old "Cluster Weighting" method).
Use the new methods as a "Sensitivity Check": If you use the new methods and get a totally different answer than the old method, it tells you that your data is tricky and the results depend heavily on how you look at it.

In short: The paper gives us better tools to check if our data is biased, but it warns us that the new tools for fixing that bias are still a work in progress. In the meantime, the simplest approach remains the most trustworthy.

Technical Summary: Estimating Association Between Paired Outcomes in Clustered Data with Informative Subgroup Size

Problem Statement
In biomedical and public health research, particularly in dental epidemiology, data often exhibit a clustered structure where observational units (e.g., teeth) are nested within clusters (e.g., patients). A critical challenge in analyzing such data is Informative Cluster Size (ICS), where the number of units within a cluster is associated with the outcome. Standard methods often fail to distinguish between participant-averaged and cluster-averaged inferences, leading to biased marginal estimates when larger clusters (e.g., patients with more teeth) are implicitly weighted more heavily.

While existing literature has addressed ICS through methods like Within-Cluster Resampling (WCR) and Cluster-Weighted Generalized Estimating Equations (CWGEE), a specific gap remains regarding Informative Subgroup Size (ISS). ISS occurs not only when total cluster size is informative but when the composition of subgroups within a cluster (e.g., the distribution of teeth across periodontal severity strata) is also associated with the outcome. Previous work has developed subgroup-weighted methods for marginal means and regression effects, but there was a lack of methodology specifically designed to estimate bivariate association (e.g., correlation between caries and periodontal disease) when both total cluster size and within-cluster subgroup composition are informative.

Methodology
The authors propose a framework for estimating measures of association (Pearson, Spearman, and Phi coefficients) between paired outcomes in clustered data under conditions of both ICS and ISS. The core of the methodology involves defining three specific weighting schemes derived from hypothetical within-cluster resampling designs. These weights are constructed by replacing random resampling estimators with their conditional expectations given the observed data.

Let $M$ be the number of clusters, $n_i$ the size of cluster $i$ , and $(X_{ij}, Y_{ij})$ the paired outcomes. The outcomes are mapped to categorical representations $K_{ij}$ and $L_{ij}$ . The proposed weights $\omega_{ij}$ are:

Population Pair Weights (PPW): Assumes all possible paired categories in the target population are relevant. The weight is inversely proportional to the number of units in the cluster sharing the specific paired category ( $n_{iP}$ ).
$\omega_{ij}^{PPW} = \frac{1}{n_{iP}}$
This targets a population where every possible pair category is equally represented, regardless of whether a specific cluster realizes all pairs.
Observed Pair Weights (OPW): Restricts inference to the distinct paired categories actually observed within a cluster. The weight is inversely proportional to the product of the number of observed paired categories in the cluster ( $N_{iP}$ ) and the count of units in that specific category ( $n_{iP}$ ).
$\omega_{ij}^{OPW} = \frac{1}{N_{iP} n_{iP}}$
This equalizes the contribution of observed paired categories within each cluster.
Marginally Observed Pair Weights (MOPW): Targets the cross-classification induced by the marginal categories observed for each outcome. It considers all combinations of observed $K$ and $L$ categories, even if some pairs are unobserved (resulting in a "null draw"). The weight is:
$\omega_{ij}^{MOPW} = \frac{1}{N_{iK} N_{iL} n_{iP}}$
This approach balances the marginal category structure without renormalizing over only the observed pairs.

To determine whether ISS is present, the authors adapt a testing procedure by Anyaso-Samuel and Datta. They employ a Stouffer's inverse-normal combination method to partition large datasets into disjoint subsets, compute subset-level rank-based statistics, and combine $p$ -values. This modification addresses the computational burden of full-sample permutation tests in large-scale dental datasets.

The variance of the estimators is calculated using a sandwich form with the delta method applied to the correlation functionals.

Simulation Results
A simulation study was conducted using a latent Gaussian model to generate clustered paired outcomes with controllable ICS and ISS mechanisms. The study evaluated the performance of the proposed weights against standard unweighted estimators and CWGEE (inverse cluster-size weighting).

ISS Testing: The Stouffer-combined ISS test performed as intended, correctly identifying informative subgroup structures when unit-level associations induced correlated subgroup membership (yielding very small $p$ -values) and remaining non-rejective in settings without such informativeness.
Estimation Performance: The results were mixed regarding the proposed subgroup weights. The standard CW (inverse cluster-size) weighting remained the most stable and reliable across simulation settings. In contrast, the PPW estimator frequently exhibited substantial bias and poor empirical coverage probabilities. The OPW and MOPW estimators showed variable performance, sometimes improving coverage but often failing to uniformly outperform CW.
Conclusion from Simulation: The proposed subgroup weights do not uniformly improve estimation of the target correlation compared to CW. Consequently, they are best interpreted as alternative marginalization schemes for sensitivity analysis rather than definitive primary estimators.

Application to NHANES Data
The methodology was applied to tooth-level data from the National Health and Nutrition Examination Survey (NHANES), analyzing the association between periodontal outcomes (Clinical Attachment Loss and Pocket Depth) and caries outcomes (Decayed and Filled Surfaces).

ISS Detection: The ISS tests revealed that subgroup-size informativeness was not uniform. Significant evidence of ISS was found for pairings involving filled surfaces (FS and FSI), likely reflecting the influence of treatment history on tooth retention. Pairings involving decayed surfaces (DS and DSI) showed little evidence of ISS.
Association Estimates: All estimated associations were positive but modest (generally below 0.20). Associations were stronger for whole-tooth caries measures compared to interproximal-only measures.
Weighting Impact: The choice of weighting scheme materially altered the estimated correlations. For filled-surface outcomes, subgroup-weighted estimates were often smaller than unweighted or CW estimates. For decayed surfaces, the direction of change was less uniform.
Interpretation: The authors conclude that the observed dentition should not be treated as an exchangeable collection of teeth. The estimated periodontal–caries association is sensitive to the marginalization scheme induced by the weighting choice.

Significance and Claims
The paper claims to fill a methodological gap by extending weighted estimation traditions (WCR, CWGEE, DWGEE) to the specific context of paired bivariate association under both ICS and ISS.

Primary Contribution: The development of three specific weights (PPW, OPW, MOPW) that allow researchers to define and estimate association under different marginal population targets when subgroup composition is informative.
Practical Utility: The introduction of the Stouffer-combined ISS test provides a computationally feasible way to detect informative subgroup structures in large-scale clustered data.
Modest Conclusion: The authors explicitly state that the simulation results do not support the proposed subgroup weights as uniformly superior replacements for CW. Instead, they position these weights as tools for sensitivity analysis. The primary recommendation is to use CW as the main estimator while utilizing the subgroup-weighted estimates to assess the robustness of conclusions to different marginalization assumptions. The paper emphasizes that in dental applications, where tooth presence is a consequence of disease history, distinguishing between primary estimation and sensitivity analysis is central to valid inference.