[Paper Review] High-recall causal discovery for autocorrelated time series with latent confounders

Andreas Gerhardus, Jakob Runge|arXiv (Cornell University)|Jul 3, 2020

Bayesian Modeling and Causal Inference35 references35 citations

TL;DR

Proposes LPCMCI, a constraint-based causal discovery method for autocorrelated time series with latent confounders, achieving higher recall than SVAR-FCI while controlling false positives. It increases effect size by using default conditioning on known parents and early orientation via novel LPCMCI-PAG rules.

ABSTRACT

We present a new method for linear and nonlinear, lagged and contemporaneous constraint-based causal discovery from observational time series in the presence of latent confounders. We show that existing causal discovery methods such as FCI and variants suffer from low recall in the autocorrelated time series case and identify low effect size of conditional independence tests as the main reason. Information-theoretical arguments show that effect size can often be increased if causal parents are included in the conditioning sets. To identify parents early on, we suggest an iterative procedure that utilizes novel orientation rules to determine ancestral relationships already during the edge removal phase. We prove that the method is order-independent, and sound and complete in the oracle case. Extensive simulation studies for different numbers of variables, time lags, sample sizes, and further cases demonstrate that our method indeed achieves much higher recall than existing methods for the case of autocorrelated continuous variables while keeping false positives at the desired level. This performance gain grows with stronger autocorrelation. At https://github.com/jakobrunge/tigramite we provide Python code for all methods involved in the simulation studies.

Motivation & Objective

• Address low recall in causal discovery for autocorrelated time series with latent confounders.
• Develop a method that increases CI test effect size by using informed conditioning sets.
• Introduce an iterative edge removal and orientation framework that determines ancestral relationships early.
• Prove soundness, completeness (oracle case), and order-independence of the proposed method.
• Provide open-source code to enable benchmarking and replication.

Proposed method

• Introduce LPCMCI, a constraint-based causal discovery algorithm for time series with latent confounders.
• Define effect size in CI testing and show how default conditioning on known parents increases power.
• Develop novel LPCMCI-PAG orientation rules and middle marks to enable early ancestor determination.
• Intertwine edge removal and orientation to iteratively refine ancestral relations during testing.
• Prove order-independence and, under oracle assumptions, soundness and completeness.
• Provide an open-source Python implementation and compare against SVAR-FCI and SVAR-RFCI baselines.

Experimental results

Research questions

• RQ1 Can LPCMCI achieve higher recall than existing time-series causal discovery methods in the presence of autocorrelation and latent confounders?
• RQ2 Does conditioning on known parents (default conditions) improve CI test power without inflating false positives?
• RQ3 Do new orientation rules and LPCMCI-PAG representations allow earlier and more reliable ancestorship determination during edge removal?
• RQ4 Is the LPCMCI algorithm order-independent and sound and complete in the oracle setting?
• RQ5 How does LPCMCI perform across varying numbers of variables, time lags, and sample sizes?

Key findings

• LPCMCI yields substantially higher recall than SVAR-FCI while maintaining false positives at the target level for autocorrelated continuous data.
• Using default conditioning on the parents of A and B increases CI test effect size and improves detection power.
• Novel orientation rules and LPCMCI-PAGs enable early determination of ancestral relationships during edge removal.
• The algorithm is order-independent and, under oracle CI decisions, sound and complete.
• Performance gains grow with stronger autocorrelation and larger problem sizes in extensive simulations.