Causation Inference in Complicated Atmospheric Environment

Ziyue Chen,Miaoqing Xu,Bingbo Gao,George Sugihara,Feixue Shen,Yanyan Cai,Anqi Li,Qi Wu,Lin Yang,Qi Yao,Xiao Chen,Jing Yang,Chenghu Zhou,Manchun Li

DOI: https://doi.org/10.1016/j.envpol.2022.119057

IF: 8.9

2022-01-01

Environmental Pollution

Abstract:Reliable attribution is crucial for understanding various climate change issues. However, complicated inner-interactions between various factors make causation inference in atmospheric environment highly challenging. Taking PM2.5-Meteorology causation, which involves a large number of non-Linear and uncertain interactions between many meteorological factors and PM2.5, as a case, we examined the performance of a series of mainstream statistical models, including Correlation Analysis (CA), Partial Correlation Analysis (PCA), Structural Equation Model (SEM), Convergent Cross Mapping (CCM), Partial Cross Mapping (PCM) and Geographical Detector (GD). From a coarse perspective, the Top 3 major meteorological factors for PM2.5 in 190 cities across China extracted using different models were generally consistent. From a strict perspective, the extracted dominant meteorological factor for PM2.5 demonstrated large model variations and shared a limited consistence. Such models as SEM and PCM, which are capable of further separating direct and indirect causation in simple systems, performed poorly to identify the direct and indirect PM2.5-Meteorology causation. The notable inconsistence denied the feasibility of employing multiple models for better causation inference in atmospheric environment. Instead, the sole use of CCM, which is advantageous in dealing with non-linear causation and removing disturbing factors, is a preferable strategy for causation inference in complicated ecosystems. Meanwhile, given the multi-direction, uncertain interactions between many variables, we should be more cautious and less ambitious on the separation of direct and indirect causation. For better causation inference in the complicated atmospheric environment, the combination of statistical models and atmospheric models, and the further exploration of Deep Neural Network can be promising strategies.