A Study on Causal Discovery Considering Confounders
Jing Song · Hokkaido University Collection of Scholarly and Academic Papers (Hokkaido University) · 2019
There are a lot of observational data in the real world in which many variables are correlated with each other. Correlation is not equal to causality. The best way to demonstrate a causal relationship between variables is to conduct a controlled ran- domized experiment. However, real-world experiments are often expensive, unethical, or even impossible. Many researchers working in various fields are thus using statis- tical methods to analyze causal relationships between variables. Many studies have been conducted to infer causality from raw observational data, but most of them have been based on the assumption that all the variables (including confounders) affecting the causal relationships have been known. Today, however, emphasis has been placed on open data. In the open data environment, it is difficult to consider all related data beforehand, and an exploratory analysis is required to acquire data that can be confounding. Therefore, in this study, first, we analyzed how the existing methods which determines the causal direction between variables are influenced by unknown confounders. Through assessing the existing methods, we found that the existing methods are susceptible to confounding in different degrees. We thus investigated how to decide whether a third variable is confounding for two observed variables. Finally, we studied on a framework to perform causal analysis while considering the possible confounders. We have three purposes for the study. Firstly, investigating a general assessment method for causal discovery methods, especially investigating their performance when the data is confounded. Secondly, investigating how to de- termine a possible common cause variable. Thirdly, investigating how to do causal analysis of open data while considering the possible confounders.