Recently I am interested in DAG and the causal inference as to Judea Pearl's work. Here is a record of my understanding.
Concepts
Front door criterion and front door path
The first thing to clarify is so-called the "Front door criterion and front door path", following is my version of understanding.
Basically, the front door criterion requires that:
- is a front door path, and in this path blocks to .
- All "back door paths" from to are blocked. Like <- A_set_of_vars -> , basically a fork like path. Here, this is indeed the case.
And basically, if we have observed a set of variables that satiesfying front door criterion, then the causal relationship between x and y is identifiable.
Back door criterion and back door path
The other concept is back door criterion and corresponding back door path. Here, we say that or are the "Z" set that meets the back door criterion relative to the causal inference from to . And we then list the conditions.
- is not the decendents of , my understanding is that there is no path like . But the reverse could be true.
- blocks all paths (regardless of the directions, as long as it connected) pointing to and also pointing or pointed at .
And backdoor criterion also provides an indentification tool.