Under what assumptions do the reported estimates have a causal interpretation?
TargetDiscovery is aimed at estimating the difference between the average outcomes in two subgroups of the sample that share the same level of the confounders but have different levels of the target variable. This difference represents the association between the target variable and the outcome adjusting for the confounders. We refer to this association as the variable importance of the target variable, to clearly distinguish it from a causal effect. TargetDiscovery uses a semiparametric regression working model that assumes a linear relationship between the target variable and the outcome, but leaves the association between the confounders and the outcome unspecified. This semiparametric regression can be written as Outcome= β Target+φ(confounders), for some unknown variable importance measure, β, and an unspecified function φ of the confounders. Note that this model assumes that there is no interaction between the target variable and the confounders. If this assumption is wrong, Targ