CEDAR Discovers Causal Edges in Sparse Autoregressive Time Series

This paper introduces CEDAR (Causal Edge Discovery for Autoregressive Processes), a novel constraint-based approach for identifying lagged causal connections within sparse autoregressive time series. The method is particularly useful when dealing with datasets where variables exhibit strong lag-1 self-dynamics and data availability is limited. CEDAR operates by first screening potential cross-variable lags using AR(1)-residualized, U-centered distance correlation. Following this, it applies two targeted conditional-independence tests for each significant candidate lag, accepting at most one lag per ordered pair. A stable MCI pruning step then removes any indirect causal edges, ensuring a clearer interpretation of direct relationships. The system also includes an optional feature for adjusting for trend-like nonstationarity using deterministic C-nodes. In sparse data environments where few lags pass the initial screening, CEDAR offers an efficient solution, requiring only O(d^2) conditional independence tests after the screening phase, while maintaining clear, edge-level interpretability. Its effectiveness is most pronounced with scarce data and dominant lag-1 self-dynamics, though other methods may be preferred with larger datasets or more complex higher-order autoregressive effects.