The basic idea is that, although correlation or statistical dependence cannot determine the causal relationship between two variables, it can, under plausible assumptions, determine some causal relationships when three or more variables are considered.