Prediction of periodontal pathology around third molars using linear mixed effects modeling

To compare the random intercept multilevel model with other linear mixed effects models in an assessment of the effect of quadrant-, jaw-, and person level covariates on probing depth of asymptomatic third molars.

Five different covariance models were considered: 1) the random intercept multilevel 2) multi-level with unequal jaw variance 3) multi-level with unequal tooth variance 4) multi-level with unequal jaw and side variance and 5) the general linear model for correlated data with unstructured covariance matrix.

235 subjects with all four third molars erupted were included. Fifty-one percent were female and 75% Caucasian. The average age was 29.1 years (sd = 7.0).

The extended multi-level with unequal residual variance was the best fit to the data. Likelihood ratio tests in a stepdown selection approach resulted in a final model for mean probing depth that included one statistically significant three-way interaction (age x gingival inflammation x gender), two statistically significant two-way interactions (jaw x gingival inflammation and jaw x gender) and one significant main effect (ethnicity).

Linear mixed effects modeling is a powerful tool for the analysis of correlated dental data. However, no one covariance structure is appropriate for all purposes.