Abstract
Randomized clinical trials are the gold standard when experimental designs are feasible. Randomization allows the handling of allocation bias for known and unknown confounders. Specific tools such as blocking, stratification, and dynamic allocation provide additional guarantees to simple randomization. When an experimental design is not feasible, the propensity score (PS) has been shown to produce greater benefit than traditional methods (i.e., restriction, stratification, matching and adjusting). There appears to be a hierarchy in terms of the effectiveness of balancing for PS techniques: matching or weighting above stratification above covariate adjustment (which is discouraged due to its drawbacks). Instrumental variable analysis and its variants might provide added value because they aim to balance for unknown confounders as well, thus mimicking randomization, but at present, are considered more useful for sensitivity rather than primary analyses.
Keywords
allocation bias - confounding - randomization - propensity scores - instrumental variable analysis
