zhihua li
2005-Mar-16 15:35 UTC
[R] Are there any R packages that can deduce causal relationship between variables?
Hi netters. Assume Y = {Y1,.....Yn} and X = {X1......Xm}, where Yi and Xi are random variables that can take on discrete values from V={0,1,2}. Each Yi in Y has some (0-k) parent variables in X, which means given the values of the parent variables (Xi0....Xik) the values of Yi is set. Yi = F(Xi0,...Xik), where F is a mapping function from parents to sons. Considering there are some noise in the data, we can put it in a probabilistic way: the parents and sons have the joint probability distribution P(Yi) = P(Yi|Xi0...Xik). Now I have a training dataset D, which includes a series of instances of Y and X. For each Yi, I want to find its parent variables (Xi0...Xik) in X and the mapping function F so that in most cases Yi = F(Xi0,...Xik). In terms of probabilistics, I want to find the joint probability distribution P(Yi|Xi0...Xik) , that best matches D. I realized it's not a simple task. I've read papers describing how to solve this problem using Bayesian Networks. But it's way too difficult for me to understand. So are there any R packages that can solve this problem in a neat way? Thanks a lot!