Bootstrapping confidence intervals in linear non-Gaussian causal model

抄録

We consider the problem of finding significant connection strengths of variables in a linear non-Gaussian causal model called LiNGAM. In our previous work, bootstrapping confidence intervals of connection strengths were simultaneously computed in order to test their statistical significance. However, such a naive approach raises the multiple comparison problem which many directed edges are likely to be falsely found significant. Therefore, in this study, we tested two multiple testing correction approaches, Bonferroni correction and Mandel's approach, then evaluated their performance. We found that both Bonferroni correction and Mandel's approach are able to eliminate some of falsely found directed edges.