On Granger causality and the effect of interventions in time series

Area of Science:

• Causal inference
• Time series analysis
• Econometrics
• Computer science
• Medical statistics

Background:

• Granger causality is widely used in econometrics for systems lacking randomized experiments, modeling dynamic development to infer causal relations.
• Intervention-based causality is prevalent in medical statistics and computer science, focusing on the effects of external manipulations.
• Existing causal inference methods often focus on single interventions or assume known causal structures.

Purpose of the Study:

• To investigate the relationship between Granger causality and the effectiveness of external interventions in multivariate time series.
• To determine conditions under which Granger causality can inform about intervention outcomes.
• To develop criteria for estimating intervention effects from observable time series, even when the full system is not known.

Main Methods:

• Combining Granger causality with intervention-based causal reasoning frameworks.
• Analyzing multivariate time series with known Granger causal structures.
• Deriving graphical criteria analogous to Pearl's back-door and front-door criteria for effect estimation.

Main Results:

• Established conditions linking Granger causality to intervention effectiveness.
• Developed graphical criteria to assess the identifiability of intervention effects from observable data.
• Demonstrated how to estimate intervention effects in systems with partial observability.

Conclusions:

• Granger causality can provide insights into intervention effectiveness under specific assumptions.
• Graphical criteria offer a practical method for determining if intervention effects are estimable from observed data.
• This work bridges two major approaches to causal reasoning in time series analysis.