"... unreliable historical control arm data."
If the historical control arm(s) data is "unreliable", what makes the contemporaneous control arm data any more reliable? It seems to me that, in effect, you indict any and all control arms as inherently, and necessarily, unreliable.
As you point out:
"Large variations in control arm are very common given that patient population makeup and previous treatment history for patients all play a role in survival in both control AND in treatment arms. That is why the control arms are added, because historical controls vary a great deal."
In other words a great number of factors, known and unknown, accompany each patient in each control arm and each treatment arm. Even the number of factors accompanying each patient is unknown. That is as true of the contemporaneous control arms as it is true of the variance present in the historical control arms.
BTW, my Masters thesis at Princeton involved Design and Analysis of Experiments in a certain 14 variable field, with all 14 variables' effect and each 2-factor interaction[but for 3] effect accurately calculated in 1032 experiments, Box-Hunter approach. G.E.P.Box was then at U. Wisconsin, and S. Hunter was still at Princeton.
You, of course, are aware of the 3 kinds of lies that I am sure even FDA statisticians are aware of: lies, damned lies and statistics. My thesis adviser took delight in constantly reminding me of the truth of that statement.
AI
Market Data 
Markets 
