Conditioning on a future (post intervention/treatment and post outcome) event in a causal diagram?

Question

My team is conducting a pre/post-intervention comparison of health outcomes in treatment and control groups, and the question came up whether it's a good idea condition/match on a deceased flag for subjects after the post study period.

Instinctively, I know contaminating your model with future information introduces bias when you're trying to predict a future event (or the causal effect of a treatment on a future event).

However, in the context of causal diagrams, how do we know conditioning on a future variable variable is not good?

Based on the diagram below, the Deceased variable is a collider which introduces unwanted bias if included by opening the Y -> D <- A backdoor path, is that correct?

Answer 1

Conditioning on a collider would indeed bias the causal estimate

I am assuming the "Diseased" variable is $D$ in your casual graph, and you are trying to estimate the causal influence $A \to Y$. In the graph, $D$ is a collider, so conditioning on it would introduce a spurious correlation between $A$ and $Y$. As you are saying, you should not condition on $D$ (you should, however, condition on the confounder $X$). This is purely a result of the do-calculus. Reasoning about the time-precedence can be helpful for coming up with a suitable graph in the first place, but it is not needed for choosing an adjustment set once the graph is set up (although it may still aid intuition).