Unified Model Selection Approach Based on Minimum Description Length Principle in Granger Causality Analysis

Fei Li, Xuewei Wang, Qiang Lin, Zhenghui HuIEEE Access 2020 · DOI

TL;DR

This paper introduces a unified approach to Granger Causality Analysis (GCA) by employing the Minimum Description Length (MDL) principle for model selection. This novel method aims to overcome the limitations of traditional two-stage GCA, which often relies on disparate mathematical theories, leading to inconsistent and potentially misleading results, especially in complex networks.

What Problem Does This Paper Address?

Granger Causality Analysis (GCA) is a powerful tool for inferring causal relationships between time series data, commonly used in neuroscience to understand brain connectivity. However, traditional GCA methods typically involve a two-stage model selection process: first, determining the model order (e.g., time lags) using criteria like AIC or BIC, and second, evaluating the significance of causal effects using F-statistics. The authors argue that this separation of model selection stages, often based on different mathematical theories, can lead to inconsistencies and a lack of robustness. Specifically, F-statistics can be sensitive to the chosen significance level and the search path through the model space, potentially resulting in spurious connections or missed true causal links. This paper aims to unify these stages under a single, principled framework to improve the reliability and objectivity of GCA.

Methods

The core contribution of this paper is the development of a unified model selection approach for GCA based on the Minimum Description Length (MDL) principle. Instead of the traditional two-stage process, the authors propose to map all candidate models within the GCA framework into a single model space where they can be compared directly based on their code length.

The MDL principle, rooted in information theory, quantifies the complexity of a model and its fit to the data by minimizing the total description length (code length). This length comprises two parts: the code length of the model itself and the code length of the data given the model. By applying MDL, the paper integrates the selection of model order (endogenous information) and the assessment of causal effects (exogenous information) into a single optimization problem.

The paper details the application of MDL in both time and frequency domains. For time-domain analysis, it formulates the problem of identifying causal influences between two time series (X and Y) by comparing the code lengths of models that include only X versus models that include both X and Y. For frequency-domain analysis, it leverages Geweke's work on spectral analysis to decompose total interdependence and identify causal influences at specific frequencies.

Datasets

The paper utilizes the following datasets for evaluation:

  • Synthetic 3-node and 5-node network experiments: These simulated datasets are generated with controlled noise levels and known ground truth connectivity, allowing for a direct comparison of the proposed method against conventional GCA.
  • Challenging fMRI dataset: This dataset involves mental calculation tasks under visual and auditory stimuli. The unique aspect is that the underlying neural correlates are expected to be similar despite different stimuli, providing a rigorous test for the robustness and consistency of the GCA methods.

Key Insights and Ideas

  • Unified Model Selection is Crucial: The paper strongly advocates for a single mathematical theory to guide all model selection steps in GCA, arguing that the traditional two-stage approach, using AIC/BIC and F-statistics, is inherently flawed due to the use of disparate theories.
  • MDL Offers Objectivity and Parsimony: The MDL principle provides a principled way to balance model fit with complexity, ensuring that the selected model is both accurate and as simple as possible, aligning with Occam's razor.
  • Improved Robustness in Complex Networks: The unified MDL approach demonstrates superior performance in identifying true causal connections and avoiding false positives, particularly in more complex network structures where traditional methods struggle.
  • Consistency Across Stimuli: In the fMRI experiment, the MDL-based GCA consistently identified similar causal networks for mental calculation tasks under different sensory inputs (visual vs. auditory), highlighting its robustness to variations in experimental design.
  • Frequency Domain Analysis Benefits from MDL: Applying MDL to frequency-domain GCA also shows promise, potentially offering a more principled way to identify causal influences at specific frequencies compared to traditional AIC/BIC-based model order selection.

Key Results

  • Synthetic Experiments: The proposed MDL-based GCA consistently outperformed conventional GCA in identifying true causal connections and reducing false positives across both 3-node and 5-node synthetic networks, especially under varying noise levels. For instance, in the 5-node network, the MDL method achieved near 100% accuracy for directly causal links and over 98.4% for other connections, while conventional GCA showed significantly lower accuracy.
  • fMRI Data: The MDL approach yielded more consistent causal connectivity results for mental calculation tasks under visual and auditory stimuli compared to conventional GCA. The similarity measure between networks obtained under different stimuli was generally higher with the MDL method, indicating greater robustness.
  • Specificity and Sensitivity: The paper suggests that the MDL principle inherently maintains high sensitivity and specificity, unlike F-statistics which can be overly sensitive to significance levels, leading to either false positives or missed true connections.

Limitations and Future Work

  • Computational Complexity: While the paper doesn't explicitly detail computational costs, the comparison between models in a unified space might introduce computational challenges, especially for very large networks.
  • Generalization to Other Function Spaces: The authors suggest that the MDL framework has the potential to accommodate Granger causality representations in other function spaces (e.g., kernel Hilbert spaces) and to compare models across these disparate spaces. This is identified as a key area for future work.
  • Adaptive Coding Schemes: Further investigation is needed to determine optimal coding schemes for MDL that are suitable for different neuroimaging modalities and noise levels.
  • Animal Experiments: The authors propose future animal experiments with simultaneous intracranial recordings and fMRI to further validate and guide the development of these MDL-based GCA forms.
  • Conditional GCA in Frequency Domain: The paper notes that conditional GCA in the frequency domain still presents some obfuscation and suggests that further work is needed to effectively distinguish direct from indirect influences in this domain.

Summary updated

BibTeX
@article{li2020unified,
  author = {Fei Li and Xuewei Wang and Qiang Lin and Zhenghui Hu},
  title = {{Unified Model Selection Approach Based on Minimum Description Length Principle in Granger Causality Analysis}},
  journal = {IEEE Access},
  year = {2020},
  doi = {10.1109/access.2020.2987033},
}
Figure