Redução da causalidade espúria em métodos paramétricos de detecção de causalidade

Abstract

Methods for detection of the cause and effect relationships between variables have been developed and are applied in many areas of knowledge. Parametric and nonparametric methods require statistical significance tests to confirm or not the existence of causality. This work focuses on the parametric methods, in which a decision that affects the tests of statistical significance is the model order. The selection of the order is made here by different criteria. Granger methods and causal relationships based on residual analysis have their performance assessed under different criteria for choosing the order of models and greater rigor in the use of statistical tests of multiple comparisons. The goal is to reduce the number of spurious causal relationships, which can produce erroneous conclusions due to errors in the existing topology between the analyzed variables. The analysis and proposed improvements are evaluated using three case studies, one with synthetic data and two with data from different industrial processes.