Currently, and apparently, data of a large study, which has been collected in the U.S. and Sweden between 1999 and 2004, is being published in two series of papers. The purpose is to report clinical and microbiological observations made in a randomized controlled trial in which seven different treatments (periodontal surgery, systemic amoxicillin and metronidazole, topical tetracycline fibers, and all combinations) were compared to standard periodontal therapy, i.e., scaling and root planing alone, in hundreds of patients with chronic periodontitis. A criticized traditional approach to data analysis where huge numbers of site-specific data were aggregated at the patient level (and highly valuable site-specific data is lost) seems to be published in our major journal, Journal of Clinical Periodontology, see, for example, the most recent paper by Socransky et al. (2013). Site-specific data analyses, at least seemingly taking into account the hierarchical structure and non-independence of observations made in a given subject are dealt with in papers which appear in the open access Journal of Oral Microbiology. I had critically commented on the previous paper by Mdala et al. (2012) [pdf] on multilevel modeling on clinical parameters after eight different treatment modalities before on this blog, see here. There I wrote,
“Unfortunately, the random parts in the models (whose design has to be considered insufficient) are not explained or even shown […] although one would like to see variance partition at the subject, tooth, and site levels, and whether common rules of thumb in decision making mainly based on site-specific clinical features hold. Frequently reported problems with extrabinomial variation, in particular underdispersion, in multilevel logistic regression models of periodontal data […] are not discussed. The main purpose of applying rather sophisticated models apparently was estimation based on just fixed effects rather than previous simple calculation of averages.”
The new paper by Mdala et al. (2013) [pdf] on bacterial counts after complex treatment of chronic periodontitis, which erroneously keeps “multilevel analysis” in the title, has abandoned the idea of multilevel modeling of bacterial data. Instead, authors turn to the more traditional approach of marginal models such as Generalized Estimating Equations (GEE). GEE, which correctly takes into account the non-independence of observations made in a given subject, may indeed be preferred when interest is mainly on the effect of explanatory variables on the response, and correlation structure is rather considered a nuisance. On the other hand, the main advantage of multilevel modeling as exercised in the previous paper by Mdala et al. (2012), with its unbiased dealing with the hierarchical structure of the data is, besides obtaining correct estimates of fixed effects, an analysis of the random part of the model, i.e., variances and covariances. It may indeed provide new and deep insights into phenomena and mechanisms operating at the level of interest, the periodontal site.
Now, authors considered three microbial responses in their negative binomial GEE models, counts of Actinomyces spp., considered beneficial; and counts of red and orange complexes as a whole. It is amazing to see that only scaling and root planing supplemented with systemic metronidazole and amoxicillin and topical tetracycline (the fibers are no longer available) was significantly superior to scaling and root planing alone at three months indicated by lower incidence rate ratios of red complex bacteria (Porphyromonas gingivalis, Tannerella forsythia and Treponema denticola). None of the other treatments significantly altered incident rate ratios of any of the microbial responses as compared to scaling and root planing alone. Moreover, after the three month examination, very few comparisons made to counts after three months (up to two years) differed significantly. Since so many comparisons were made using the data set they might be regarded spurious.
The authors conclude that (and further explain results presented in supplementary material which I wasn’t able to access),
“[S]hort-term reduction in counts of the red complex in deeper sites were observed with AMOX+MET+TET. However, the treatments did not produce significant beneficial changes in counts of the orange complex and Actinomyces in these sites. After treatment red and orange complex counts significantly increased in smokers, and smoking significantly reduced counts of Actinomyces. Count levels of the orange and green complexes were significantly lower in Swedish subjects, while Actinomyces counts were significantly higher compared to American subjects. Further increase in PD significantly elevated counts of the red and orange complexes. Counts of the red complex were also higher in sites that were bleeding, lost more attachment, and had gingival redness.
We found that BOP, accumulation of plaque, deeper pockets, and smoking had detrimental effects on the counts of the complexes and believe that these four factors were mainly responsible for diminishing the effects of the study treatments. Our study clearly showed that antibiotic-treated patients tend to suffer relapse of the microbiota associated with periodontitis unless dental plaque is prevented from re-accumulating either by self-inflicted personal oral hygiene or professional maintenance therapy. The importance of monitoring risk predictors during maintenance therapy such as smoking and probing depth was also indicated.”
Few of the second paragraph should be accepted at face value. There might be circumstantial evidence that quality of maintenance (and smoking status) rather than choice of treatment modality dictates long-term results, but that we know for decades.
Considering bacterial complexes as response variables and clinical parameters as independent covariates in statistical models is a rather new way of thinking, at least for these well-known scientists. So far, they had tried to convince the reader that bacteria, in particular the red and orange complexes, cause the disease. Now, these authors seem to regard colonization a corollary for the first time. They might indeed be correct.
What would be more interesting, say for instance, in multilevel logistic regression models considering deep periodontal lesions, is successful resolution of periodontitis at the site level as response. Different therapies, plaque levels, smoking etc., and certain bacterial complexes can then be entered as lower and higher covariates. In particular the random part could reveal new insight into subject level and tooth level variation which has, so far, rarely been described.
9 October 2013 @ 1:10 pm.
Last modified October 9, 2013.