Multilevel Models For Gingival Bleeding in a Steady-state Plaque Environment

Objectives: Models fitting longitudinal bleeding on probing were to be explored in order to develop a multilevel representation of a steady-state plaque environment in plaque-induced gingivitis. Material and Methods: 50 non-smokers (16 male) with mild gingivitis (19 and 28 yr old), participated. Periodontal conditions were recorded at 6 sites of every tooth present. Probing parameters were measured to the nearest mm, and bleeding on probing assessed with a pressure-controlled probe. The amount of plaque was estimated using criteria of the plaque index (scores 0-3) system. Examinations were repeated after 2 and 4 weeks while volunteers were exhorted not to change oral hygiene habits. We modeled the binary response, bleeding on probing. Fit and properties of a standard univariate, 4-level (occasion, site, tooth, subject), repeated measures model, and multivariate multilevel models for repeated measures were explored. Results: Model assumptions of the univariate repeated measures model were generally violated leading to unstable parameter estimates. The multivariate multilevel binary response model revealed a better fit of the data. Omitting the lower (tooth) level from the analysis virtually removed extra-binomial variation, whereas including site, tooth, and subject-related covariates in the model did not. Bleeding was strongly related to papilla with odds ratios of between 1.52 (95% CI of 1.32-1.75) and 1.68, (1.47-1.93). With each increase in plaque index score, the odds of bleeding increased by 29 up to 41%. Similar increases of between 41 and 54% were noted with each increase of periodontal probing depth by 1 mm. Conclusions: The present study reveals that the correct choice of the multilevel model is crucial for the interpretation of fixed and random parameter estimates, in particular if longitudinal binary response data are collected in periodontal trials.