Introduction
In macroeconometrics, Vector Autoregressive (VAR) and Dynamic Factor (DF) models are standard tools for forecasting and summarizing economic data. However, these models typically yield "reduced-form" results that capture statistical associations rather than underlying structural (causal) relationships. This creates a significant challenge for policy analysis: to predict the outcome of an intervention, such as a change in interest rates, researchers must identify the structural shocks. Traditionally, this identification relies on strong, often subjective, theoretical assumptions. This paper presents a semi-automated search procedure using graphical causal models to discover these structural restrictions directly from the data's statistical properties.
Methodology: Graphical Causal Discovery
The authors adapt causal discovery algorithms—specifically the PC, FCI, and CCD algorithms—to the time series domain. The process begins by estimating a standard VAR or DF model to extract reduced-form innovation terms (residuals).
Assuming the Causal Markov and Faithfulness conditions, the methodology employs a systematic search:
- Initial Graph: Start with a fully connected undirected graph of contemporaneous variables.
- Edge Elimination: Use vanishing partial correlations to delete edges where variables are conditionally independent.
- Directionality: Direct the remaining edges to form a Directed Acyclic Graph (DAG) or, in more complex cases involving feedback loops and latent variables, an undirected graph that serves as a filter for possible structures.
Empirical Application: US Monetary Policy
The approach was validated using US macroeconomic datasets, including those from Bernanke and Mihov (1998). The variables analyzed include GDP, the GDP deflator (PGDP), commodity prices (PSCCOM), the federal funds rate (FFR), total bank reserves (TR), and nonborrowed reserves (NBR).
The algorithm's output for these variables (as seen in the figure above) provides a data-driven map of the contemporaneous causal structure of the US central bank's operating procedures. By combining this graphical output with minimal background knowledge, the authors isolated the exogenous monetary policy shock—specifically identifying it as the portion of the shock to nonborrowed reserves (NBR) orthogonal to total reserves (TR).
Key Results and Insights
The data-driven identification process yielded several significant findings:
- Challenging the Real Business Cycle Hypothesis: Shocks to nominal variables (money and inflation) were found to have substantial causal effects on real macroeconomic fluctuations, contradicting theories that only productivity shocks drive output.
- Impulse Response Validation: To verify the identified structure, the authors generated impulse response functions (IRFs). The figure below illustrates the response of GDP to a one-standard-deviation monetary shock over 48 months. The resulting dynamics are coherent and align with economic intuition, demonstrating the validity of the graphical search procedure.
Limitations and Future Directions
While powerful, the methodology faces challenges in distinguishing between feedback loops and latent unmeasured variables using only partial correlations. In such cases, the algorithm may output undirected edges, requiring external domain knowledge or additional statistical tests for full identification. Future research aims to refine these algorithms to better infer the properties of directed cyclic graphs in the presence of latent confounders.
Summary updated