C.6 Bugs Fixed in Version 3.2.2 of ChemTree
- When a split was done on a continuous response, and all daughters had zero variance, the results were inconsistent. Previously, if there were small roundoff errors in detecting that each daughter had zero variance, two different results were possible depending on the roundoff errors. One is, the p-value was computed as exp(-69), the second was that the split was dropped from consideration. Theoretically, such splits should be impossible under normal distribution assumptions (zero probability), but they still happen in practice. Now a fix has been put in place to use the variance from the root node as an estimate for the variance within such splits, and treat them consistently, irregardless of roundoff errors. As a consequence of this fix, trees generated with this new release will potentially differ from those created in previous releases.
- (Release 3.2.3:) The above bug fix introduced a new bug whereby when a split was attempted on a multivariate binary response, with a categorical predictor, the split was not always found. This has been fixed.
- If a cluster plot (observation distance matrix plot) was done from one of the branch or leaf nodes of a tree, and another tree is made from a rubberband on the cluster plot, the statistics on the second tree used to be wrong. This has been fixed.
- Increased reliability of using more than one parallel thread for splitting when the computer has more than one processor.
- Removed "Nodename" header for "Save Average Tree Predictions" output.