Technically, 'taxonomy 3' is an Eigen decomposition based on correlations or covariations of contrasts of empirically-derived log Bayes factors.

The main features of the method can be tested on-line with small datasets: see the on-line analysis page.

The method can be divided into the following main steps:

• Calculation of Log Bayes Factors (LBF). The variables measured in cases and controls (e.g. categorical variables such as genotypes) are transformed into a new continuous measure: 'LBF'. This measure amplifies each individual's contribution to the overall between group distinction (i.e. case/control distinction), and produces variable-based discriminatory or classificatory evidence. LBFs are a subset of the divergence mathematical group and have essential properties allowing the following steps.
  
  
• Aggregation of LBFs (optional). LBFs can be regrouped (e.g. averaged) before being further analyzed. The grouping can use groups of variables already known (biologically meaningful: genes, ontologies). This allows the ‘supervised’ biological exploration of data, and the classification of subjects (sub-phenotype).
  
  
• Principal component analysis. Patterns of co-occurence between LBFs are looked for. Variables in correlation with the case/control distinction (the between group signal) are separated from i) variables in correlation with any within group signal (i.e. population heterogeneity or population variability observed only in cases, or only in controls) and from ii) the remaining variables (noise).
  

The method can deal with any type of variables or outcomes (e.g categorical or continuous), the software we provide can deal with categorical outcomes (e.g. case/control studies) and variables of many data types (genetic, discrete, continuous : normal, exponential, poisson).

 

top of page