Clinical measurement of periodontitis has historically focused on the concept of periodontitis as a slow, continuous process which has emphasized measurements of the static condition of periodontal pockets. Observations based on longitudinal measurement of attachment loss in untreated subjects have indicated that periodontal destruction occurs in discrete episodes of short duration. Based on these studies, it has been suggested that chronic periodontal disease proceeds through a series of random episodic attacks. Periodontal sites are considered as existing in 2 states, either disease active or inactive. During periods of disease activity, sites increase in their probable depth, whereas during the inactive state, no significant change in probing depth can be detected. The detection of changes at periodontal sites from time series data has been addressed by 3 analytical procedures: regression, running medians, and tolerance. The standard deviation of differences between replicate measurements of 48,064 sites for 56 subjects was 0.7727 mm. From this estimate, the computed standard deviation for a single measurement was 0.5464 mm and for the mean of 2 measurements was 0.386 mm. The expected error rates of each method have been estimated by computer simulation. The type-I error for the regression (p = 0.028), running median (p = 0.000025), and tolerance (p = 0.00012) methods were all sufficiently low to consider it unlikely that reported observations could be accounted for by methodologic error. The estimated type-II error for the regression (p = 0.446), running median (p = 0.152), and tolerance (p = 0.068) methods suggests that a substantial fraction of disease active sites was not detected by these methods. Several data set properties have been investigated. Intraclass correlation coefficients were computed from attachment level changes on 8,130 sites in 105 patients. By this analysis, 7% of the variation was associated with the subject and 93.3% with the individual sites, indicating that attachment level changes at periodontal sites exhibit a high degree of statistical independence. Autocorrelation within sequential attachment level measurements was computed and found low (0.081 in 22 subjects and 0.099 in 45 subjects), indicating that computed variance is not systematically underestimated due to autocorrelation within the data set. Clinical measurements which have failed to exhibit association with episodic attachment loss include gingival redness, bleeding on probing, suppuration, supragingival plaque, and darkfield microscopic bacterial counts.(ABSTRACT TRUNCATED AT 400 WORDS)