The causal analysis in the log-linear model
Gloria Gheno · Padua@research (University of Padova) · 2015
In this paper I try to provide a causal theory which can be applied to the log-linear models because they are devoid of complete methods to calculate the different effects. My causal theory uses odds ratios and Pearl's definitions (2001, 2009, 2012). I find that although I delete the log-linear parameter which regulates the interaction, there is still an interaction effect, which I call cell effect and which is due to the mere presence of two variables which affect a third. I then try to calculate this effect using Pearl's formulas for the direct, indirect and total effects and I find that a part of this effect is contained in Pearl's direct natural effect, this part is that related to the nonlinearity. To conclude I provide a test to see if this effect is equal to 0.