Fig. 1

Causal diagram showing conditions that will result in survival bias in an RCT. R = treatment, S = binary intermediate response (e.g. pregnancy, live birth), U = uncontrolled or unmeasured baseline variables, and Y = study outcome. Arrows depict causal relationships. For parsimony of presentation, we label the paths with the corresponding parameters from the data generating model used in a simulation study. A box drawn around S = 1 indicates that we are conditioning on the intermediate variable: the analysis set includes only participants with S = 1. Bias will occur whenever α_R, α_U, and β_U are all nonzero