In studies that use observational data, the confounding of the treatment and the covariates (observed and unobserved) can threaten their scientific validity. For such studies and for historically controlled trials, it can be helpful at the design stage to conceptualize the target randomized design one would have designed to assess the comparative effectiveness and then plan the use of non-randomized data accordingly. The prospective use of propensity scores can address the confounding of the treatment with the observed covariates but not the unobserved ones. Other limitations of propensity scores for regulatory decision-making are discussed. Even in randomized trials, non-adherence and loss-to-follow-up can reduce the study to an observational one for which confounding needs to be addressed, one way of which is using principal stratification. All these methods can help to address causality in the presence of confounding.