Causality from Observational Data (303088)

Keywords: kernel regression, stochastic dominance, exogeneity, additive noise models

Causality is one of the most important issues in all of science. Many social sciences do not have the luxury of using expensive, time-consuming and sometimes unethical controlled experiments to determine the causal direction. We evaluate two criteria choosing the direction XàY if the kernel regression of Y on X provides better forecasts or is more robust than the opposite regression of X on Y. A third criterion chooses XàY if suitably defined "virtual manipulation" of stimulus and response using kernel gradients indicates a stronger effect of X on Y, than vice versa. Simplified versions of my stochastic dominance algorithms (originally meant to choose among investment portfolios) quantify the causality criteria. This presentation will describe a matrix of generalized correlation coefficients revealing the causality of great potential use in Big Data and machine learning. I explain and illustrate R algorithms for implementing the three causality criteria by using simulations showing impressive success rates and real world examples and multivariate extensions.