One cannot use observational data to estimate within a decision analysis, without further assumptions.

Imagine =heart attack, =drug A, =smokes, and that smoking leads one to avoid drug A and get more heart attacks.

If we use observational data, but do not collect smoking status, it would appear that drug A reduces heart attacks, even if drug A truly has no effect.

I suspect one needs to treat like a counterfactual in some form, and therefore that one would need to rely on the no unmeasured confounders assumption, if one is using observational data.