Discovering Causal Relationships Between Time Series With Spatial Structure

1. Problem Statement

The paper addresses a critical gap in causal discovery (estimating the structure of cause-and-effect relationships) regarding spatiotemporal data.

• The Challenge: Existing causal discovery algorithms, particularly those designed for time series (e.g., PCMCI+), assume data is independent and identically distributed (iid) or handle only temporal autocorrelation. They fail when applied to data with spatial structure because:
  • They ignore spatial autocorrelation (where nearby locations influence each other).
  • They cannot handle latent spatial confounders (unobserved variables with spatial structure that distort causal relationships).
  • Current spatiotemporal methods often require summarizing data across locations (losing resolution) or scale poorly to high dimensions.
• The Goal: To extend time-series causal discovery to systems with spatial structure, specifically targeting "latent mechanism" discovery. Unlike "latent pattern" discovery (which focuses on spatiotemporal trends), latent mechanism discovery aims to infer the underlying causal relationships between variables that hold true across space and time, which is vital for fields like ecology, public health, and environmental science.

2. Methodology

The authors propose a framework that integrates constraint-based causal discovery with methods for spatial confounding adjustment.

Core Algorithmic Logic

The framework builds upon the PCMCI+ algorithm (a constraint-based method for time series) and adapts it using the Joint Causal Inference (J-PCMCI+) logic:

1. Contextualization: Spatial locations are treated as "contexts" (exogenous variables).
2. Separating Sets: In standard PC algorithms, a separating set $A_{ij}$ is found to test if $X_i \perp\!\!\perp X_j | A_{ij}$. The authors propose including spatial contexts (coordinates or neighborhood structures) by default in these separating sets.
3. Deconfounding: Since spatial contexts are exogenous (they do not change based on the system variables) and cannot be colliders, their inclusion in separating sets prevents false negatives in independence tests caused by latent spatial confounding.

Technical Implementation

• Conditional Independence Testing: The framework utilizes the Generalized Covariance Measure (GCM) test (Shah & Peters, 2020).
  • Instead of linear regression, the GCM test approximates conditional distributions using flexible non-parametric models.
  • Model Choice: The authors recommend Penalized Generalized Additive Models (GAMs) with spatial smooths (e.g., thin-plate splines).
  • Rationale: GAMs can model complex spatial dependencies and meet the Out-of-Sample Mean Squared Prediction Error (O-MSPE) criteria required for GCM asymptotic power, without the data-splitting issues inherent in tree-based or kernel methods when applied to spatial data.
• Algorithm Flow:
  1. Lagged Phase: Identifies causal links between past and present variables (standard PCMCI+ logic).
  2. Instantaneous Phase: Identifies simultaneous causal links. Here, the search for separating sets explicitly includes spatial coordinates/neighborhoods to control for spatial confounding.
  3. Orientation: Edges are directed based on temporal precedence (lagged) and collider rules, avoiding feedback loops.

3. Key Contributions

• Framework Extension: The paper proposes the first theoretical framework to adapt time-series causal discovery (specifically PCMCI+) for spatially structured time series to uncover latent mechanisms.
• Integration of Spatial Confounding: It bridges the gap between causal discovery and causal effect estimation by applying techniques used to handle latent spatial confounders (using coordinates as proxies) directly into the independence testing phase of discovery.
• Methodological Guidance:
  • Recommends Penalized GAMs over other machine learning models for the GCM test in spatial settings to preserve spatial information and ensure smoothness.
  • Discusses the trade-offs of knot selection in GAMs for this specific purpose.
• Critical Analysis of Assumptions: The authors rigorously analyze the necessary assumptions:
  • Exogeneity: Spatial contexts must not be influenced by the system variables.
  • Faithfulness: Discusses how PC-based algorithms search for minimum separating sets and potential violations.
  • Sufficiency: Acknowledges that while spatial confounders are handled, non-spatial latent confounders remain a challenge.
  • Stability: Assumes relationships are stationary over time, though notes potential extensions via "regimes."

4. Results and Evaluation

Note: As this is a "short communication" outlining a developing framework, it does not present final empirical results or benchmark performance metrics.

• Theoretical Validation: The paper argues that the proposed logic is sound based on existing theoretical results in causal discovery across contexts and spatial causal inference.
• Simulation Strategy: The authors propose using Agent-Based Models (ABMs) rather than standard additive noise hierarchical models for future benchmarking. They argue ABMs better simulate the mechanistic rules of real-world systems and avoid the "variance inflation" issues found in standard simulations that bias causal discovery algorithms.
• Performance Metrics: The paper critiques current metrics (like F1 score) for causal discovery, suggesting that Structural Intervention Distance or qualitative comparisons of graph interpretation are more appropriate for spatiotemporal data, as some errors (e.g., missing a mechanism vs. misidentifying timing) have different consequences.

5. Significance and Future Directions

• Scientific Impact: This framework enables researchers in ecology, epidemiology, and economics to move beyond descriptive spatiotemporal patterns to mechanistic understanding. It allows for the empirical derivation of causal models from complex observational data where randomized experiments are impossible.
• Policy Relevance: The resulting causal graphs can be translated into Qualitative Influence Diagrams (QIDs), which are widely used by agencies (e.g., US EPA) for environmental policymaking. This bridges the gap between statistical rigor and decision-making.
• Future Work:
  • Formalizing the algorithm and running extensive simulations using ABMs.
  • Developing uncertainty quantification methods (e.g., bootstrap aggregation adapted for spatial data) to assess the confidence of edge assignments.
  • Addressing computational intensity through parallelization strategies (e.g., parallel-PC) while managing the interdependence of causal discovery stages.
  • Investigating non-stationarity (changing relationships over time) within the spatial framework.

In summary, the paper provides a foundational theoretical and methodological blueprint for solving the "spatial confounding" problem in causal discovery, offering a path toward more robust, interpretable, and actionable causal models for spatiotemporal data.