It is also not easy if you have many potential covariates! Because statistically, you want a complete (explaining all effects) but parsimonious (using as few pr

It is also not easy if you have many potential covariates! Because statistically, you want a complete (explaining all effects) but parsimonious (using as few predictors as possible) model. Yet you by definition don‘t know the true underlying causal structure. So one needs to guess which covariates are useful. There are also no statistical tools that can, given your data, explain whether the model sufficiently explains the causal phenomenon, because statistics cannot tell you about potentially missing confounders. A cool, interesting, horrible problem to have :)