Causation inference in complicated atmospheric environment.

Area of Science:

• Environmental Science
• Atmospheric Science
• Data Science

Background:

• Understanding climate change impacts requires reliable attribution of causes.
• The atmospheric environment presents complex, non-linear interactions, challenging traditional causation inference.

Purpose of the Study:

• To evaluate the performance of various statistical models in inferring causation between PM2.5 and meteorological factors.
• To identify the most effective methods for disentangling complex interactions in atmospheric systems.

Main Methods:

• Comparative analysis of six statistical models: Correlation Analysis (CA), Partial Correlation Analysis (PCA), Structural Equation Model (SEM), Convergent Cross Mapping (CCM), Partial Cross Mapping (PCM), and Geographical Detector (GD).
• Application of these models to PM2.5-meteorology data from 190 cities in China.

Main Results:

• Top meteorological factors influencing PM2.5 showed general consistency across models at a coarse level.
• Significant model variations and limited consistency were observed when identifying the dominant meteorological factor.
• Models like SEM and PCM struggled to accurately separate direct and indirect causation.
• Convergent Cross Mapping (CCM) demonstrated superior performance in handling non-linear causation and confounding factors.

Conclusions:

• Sole reliance on multiple statistical models for causation inference in complex atmospheric environments is not feasible due to inconsistencies.
• Convergent Cross Mapping (CCM) is a preferred strategy for inferring causation in complicated ecosystems.
• Caution is advised when separating direct and indirect causation due to multi-directional and uncertain interactions.
• Combining statistical and atmospheric models, alongside exploring Deep Neural Networks, offers promising future strategies for improved causation inference.