Closed ghost closed 1 year ago
In this DAG, adjustment on X2 creates what is called "irremediable bias". This is different from a back-door path (because the biasing path does not start with an arrow into X), and cannot be redeemed anymore by conditioning on X1. The reason is that even after conditioning on X1, you still have the open "path" T -> Y -> X2 <- Y <- T -- this is not technically a path, as it contains nodes twice, but it still correctly visualizes the source of the bias in this case.
I have made the following model in DAGitty:
Where X2 is controlled for.
DAGitty says:
My understanding is that I can estimate the total effect if I control also for X1, other than X2, as that would block the backdoor path opened by controlling for X2.
Is that wrong?