Accounting for Intra-class Correlations (ICC) in Practice-Based Dental Research

Objectives: The intraclass correlation coefficient (ICC) is a measure of pairwise correlation between observations within the same group. Although numerous studies have published ICC estimates for medical outcomes, little is known about ICC between patients within dental practices. Accounting for ICC is important for preserving inference in analysis of group-randomized trials and community-based research, where presence of ICC can substantially impact the precision of estimates. Also, while analysis can account for ICC without explicitly estimating it, reasonable ICC estimates are crucial to sample size calculations when planning cluster-based trials. Our main objective is estimating ICC for key outcomes measured in PRECEDENT network trials.

Methods: Cross-sectional data on 1943 patients (ages 2-93, 45% male) were obtained from the first PRECEDENT trial, which captured dental outcomes at 101 dental practices in the northwest United States. We estimated ICC for various treatment-pattern and disease outcomes, using generalized estimating equations (GEE) with exchangeable working covariance.  We then investigated how these estimates change after adjustment for practice-level and patient-level covariates in the GEE regression framework.

Results: Our initial results show a tendency for variables that are more susceptible to dentist influence, such as use of prophylaxis or treatment type, to have high ICC (>0.11). Use of fluoride had especially high ICC (>0.5). Variables that are more patient-dependent, such as presence of disease, have comparatively low correlation (~0.05 to 0.1); however, even these relatively low ICC values may substantially impact statistical inference. ICC is attenuated slightly by adjustment for some practice-level covariates, such as DDS experience level, practice type, and state of residence, and to a lesser degree by adjustment for patient-level characteristics.

Conclusion:  Higher ICC for variables that involve treatment choice may reflect differences in treatment approaches among dental practices. Correlation within dental practices can be quite high and have a large impact on inference.