Revisiting Causality Inference In Markov Chain

Identifying causal relationships is a key premise of scientific research. Given the mass of observational data in many disciplines, new machine learning methods offer the possibility of using an empirical approach to identifying unappreciated causal relationships and to understanding causal behavior. Conventional methods of causality inference from observational data require a considerable length of time series data to capture cause and effect relationships. We believe that important causal relationships can be inferred from the composition of one-step transition rates (Markov Chains) to and from an event. Here we introduce 'Causality Inference using Composition of Transitions' (CICT), a computationally efficient method that reveals causal structure with high accuracy. We characterize the differences in causes, effects, and random events in the composition of their inputs and outputs. To demonstrate our method, we have used an administrative inpatient healthcare dataset to set up a graph network of patients transition between different diagnoses. Then we apply our method to patients transition graph, revealing deep and complex causal structure between clinical conditions. Our method is highly accurate in predicting whether a transition in a Markov chain is causal or random and performs well in identifying the direction of causality in bidirectional associations. Moreover, CICT brings in new information that enables unsupervised clustering methods to discriminate causality from randomness. Comprehensive performance analysis using C-statistics, goodness-of-fit statistics and decision analysis of predictive models, as well as comparison with the medical ground truth, validates our findings.