When causality matters for prediction: Investigating the practical tradeoffs
Robert E. Tillman and Peter Spites (Carnegie Mellon University, Pennsylvania)
Recent evaluations have indicated that methods used to predict the value
of a target variable after predictor variables are manipulated, e.g. when
training data is from an unmanipulated population and test data from a manipulated
population, which ignore causality actually perform better in some cases
than well known methods for causal discovery. In light of these results,
we investigate the total contribution to prediction error made when non-causal
methods use incorrect predictors for a manipulated distribution and when
causal methods use incorrect or biased parametric constraints. We give theoretical
conditions, which we experimentally confirm, for manipulations where causal
methods for prediction should have no advantage over non-causal methods and
for manipulations where causal methods should produce considerably fewer
errors, which may help to explain these surprising results.