Title: A transfer entropy test for causality in longitudinal data
Authors: Andres Romeu - Universidad de Murcia (Spain) [presenting]
Maximo Camacho - Universidad de Murcia (Spain)
Manuel Ruiz-Marin - Universidad Politecnica de Cartagena (Spain)
Abstract: Granger causality (GC) is the most popular statistical definition of causal relationship. In this type of analysis avoiding sample selection problems often requires data from many individuals and many time periods, whose observations are pooled in cross section or collected in panel data. However, this testing procedure could give a misleading account of the causality effect due to typical data problems. In particular, the size and power of this test can be seriously affected when the linearity assumption breaks down, such as under random coefficients, heterogeneous panels, structural breaks, and extreme observations. We propose a causality test based on the concept of transfer entropy between two variables that is simultaneously robust against these data problems. The test uses the pooled sample to disclose causality between the variables of interest of a generic form without having to impose an ex-ante structure of heterogeneity in the causal relationship. A Monte Carlo experiment with five different scenarios shows that the test displays correct size and high power to detect causality in situations where linear GC fails. We provide two empirical examples. One uses per capita GDP and government expenditure yearly World Bank data in a panel of 100 countries for the 1961-2016 period. Second, we test causality between firm size and productivity using data from the Bureau of Labor Statistics on 86 manufacturing firms for the 1988-2015 period.