Fig. 3: Graphical representation of the leave-one-study-out resampling algorithm.
From: The design and evaluation of hybrid controlled trials that leverage external data and randomization
Step (i), we randomly sample with replacement n patient profiles and the corresponding outcomes from the control arm (SOC) of study k. Step (ii), we use the control arms of the remaining studies as externally controlled (EC) data. Step (iii) we randomize n1 of the patients in Step (i) to the experimental treatment (EXPT) and the SOC arms of our in silico trial and compute the index \({{{{{{\rm{W}}}}}}}_{1}.\) If \({{{{{{\rm{W}}}}}}}_{1}\le {{{{{{\rm{w}}}}}}}_{1}\)(\({{{{{{\rm{W}}}}}}}_{1} \, > \, {{{{{{\rm{w}}}}}}}_{1}\)), the futility interim analysis (IA) leverages (does not leverage) EC data, and we use the ratio r2,C:r2,E (r1,C:r1,E) for the remaining n2 = n−n1 patients during the 2nd stage. For the final analysis, we recompute the dissimilarity index \({{{{{{\rm{W}}}}}}}_{2}\), and use (don’t use) EC data for inference on treatment effects if \({{{{{{\rm{W}}}}}}}_{2}\le {{{{{{\rm{w}}}}}}}_{2}\)(\({{{{{{\rm{W}}}}}}}_{2} \, > \, {{{{{{\rm{w}}}}}}}_{2}\)). We repeated these Steps (i) to (iii) 2000 times using different random samples.
