Assuming a two-sided two sample t test with a significance level (α) of 0.05 and a power (β) of 80% required a sample size of 34 implants. A 20% compensation for dropouts was taken into account (34/0.8 = 42.5 implants). Based on a previous study [10], it was expected that not all baseline microbiological samples would yield a detectable number of cultivable bacteria ([10], 19 out of 79 = 24% of samples showed no bacterial growth). Because “negative” samples cannot be used to determine a decontaminating effect, the sample size was compensated for these potential unusable samples (24%), yielding a sample size of 56 implants (42.5/0.76). According to the assumption that each patient has on average more than two implants with peri-implantitis [10], a sample size of 28 patients was chosen (56/2, 14 patients per group).

Statistical analysis

For the analysis of the primary outcome variable and the secondary microbiological outcome variable linear regression analysis was performed. The implant was taken as the statistical unit. Total anaerobic bacterial loads at baseline (Tpre and T0) were distributed normally after logarithmic transformation. Baseline values were included in the regression model. For the comparison of the number of culture-positive implants after the decontamination period, the chi-square test was used. The secondary clinical outcome variables were analyzed using a two-level hierarchical random intercepts model. The two levels of analysis were implant level and patient level. With the crude analysis, the effect of the intervention was determined, while controlling for baseline value. Because a previous study [9] has shown that mean bone loss at baseline and smoking are prognostic indicators for the outcome of resective peri-implantitis treatment, these factors were additionally included in the model (adjusted analysis).

Descriptive data and data regarding the microbiological outcome variables were analyzed using IBM SPSS Statistics 22 Version 22.0 (IBM Corp. Armonk, NY: IBM Corp.). Multilevel models were analyzed using MLwiN version 2.12 (Centre for Multilevel Modeling, University of Bristol, Bristol, UK).