Clustered dental data commonly arise when multiple teeth or tooth sites are observed within the same individual. In such settings, the number of observed units within a cluster may be informative, since tooth loss and missing measurements often reflect underlying oral health status. Standard marginal association measures may therefore be biased when larger or smaller clusters contribute disproportionate information. This paper develops weighted estimators for marginal association between paired tooth-level outcomes in the presence of informative cluster size and informative within-cluster subgroup structure. The proposed approach extends the logic of within-cluster resampling and cluster-weighted estimating equations to paired bivariate outcomes by constructing weights that balance contributions across clusters, observed marginal categories, and observed paired categories. Weighted estimating equations are used to estimate moment, rank, and cell-probability functionals, yielding clustered-data analogues of Pearson, Spearman, and phi association measures. Sandwich variance estimators and delta-method standard errors are derived for inference. Simulation studies assess finite-sample bias, standard error estimation, and coverage under varying sources of cluster-level and unit-level dependence, as well as outcome-dependent observation mechanisms. The methods are illustrated using tooth-level periodontal and caries outcomes from NHANES, where informative subgroup-size diagnostics indicate that the observed distribution of disease severity is not independent of within-mouth structure. The proposed estimators provide a principled basis for estimating marginal oral-health associations for a typical tooth from a typical individual, while reducing bias induced by informative tooth retention and subgroup composition.
🌊
