How is node cost measured in the randomForestSRC multivariate regression case? Is it by the Euclidean distance or by the Mahalanobis distance (as with the MultivariateRandomForest package)?
randomForestSRC does not use MD, which only applies to continuous settings. We use a composite univariate splitting rule thus allowing us to handle mixed outcome regression settings (ie. when you have mixture of categorical and ordinal Y values).
The composite rule is an average of the individual outcome splitting rules, where each of the outcomes have been standardized so as to allow an equal weighting.
For more information, you can see Section 8 in the online documentation:
https://kogalur.github.io/randomForestSRC/theory.html#section8