The Dartmouth CGL
is pleased to announce the release of version 0.6.2 BETA of our open-source multifactor dimensionality reduction
(MDR) software package. The new version can be downloaded from Sourceforge.net
by clicking here
. The new version is able to model datasets that have more of one class than the other (e.g. more controls than cases). MDR models interactions in imbalanced datasets using 'balanced accuracy' in combination with a threshold for assigning risk that is equal to the ratio of class 1 (e.g. cases) to class 0 (e.g. controls) in the data. The use of balanced accuracy replaces the old metric based on just accuracy. Here, balanced accuracy is the arithmetic mean of sensitivity and specificity or (sensitivity+specificity)/2. We have a paper that is in prepration that shows using simulated data that balanced accuracy is superior to accuracy when the data are imbalanced. Of course, when the data are balanced, the two metrics are identical. The MDR permutation testing module has also been updated to version 0.4.5.
The next release of open-source MDR will include new graphical output to facilitate the statistical interpretation of MDR models. A future version will also include the interactive ability to construct attributes using the MDR algorithm for inclusion in the dataset and reanalysis. Our paper that is 'in press' in the Journal of Theoretical Biology
(see post from Nov. 30, 2005) describes some of the ideas behind these new features.
We hope to have MDR 1.0 ready by the end of Feb. Please send us your comments, suggestions, and criticisms on the BETA version before then.