Simulation study of methods to detect periodontal associations when they are inconsistent among subjects.

Most statistical methods used to evaluate associations between indices of clinical periodontal diseases and purported prognostic markers test for effects across subjects. If associations exist within only a subset of subjects, however, associations may be masked, particularly in small studies. This issue was explored by using simulation to study four methods for detecting periodontal associations. Built into the simulations was the possible biological reality that a non-zero association between the two variables of interest (squared correlation coefficients, rho 2, ranged from 0.1 to 0.9 depending on simulation), measured at 16 sites per subject, did not exist in all of 10 hypothetical subjects. The four methods for testing the null hypothesis that rho = 0, or a related hypothesis; were: (1) Sites, analysis based on 160 sites incorrectly considered independent observations; (2) Subjects, analysis based on one score for each of 10 subjects; (3) Each subject, separate analyses based on sites within each of 10 subjects, family-wise type I (alpha) error corrected for multiplicity, and (4) the Each Subject method where P-levels were estimated using permutation procedures rather than t-distributions. Each Subject methods were found to have greater relative power (although there are differences in null hypotheses) under conditions of heterogeneity in rho and are considered to be particularly relevant in exploratory periodontal research when the primary interest is establishing the existence of a relationship, even if in only a subset of subjects.