ABSTRACT
I. Introduction.......................................................................................... 231
II. Bias ...................................................................................................... 232
III. Modifications Using “Blinded” Data .................................................. 234
A. Blinding......................................................................................... 233
B. Partial Unblinding......................................................................... 234
C. The Null Neutral Principle ........................................................... 235
D. Adaptive Statistical Analysis Planning ........................................ 237
E. Scope and Limitations .................................................................. 237
IV. Modifications Using Unblinded Data.................................................. 237
A. Bias and Naive Analysis............................................................... 237
B. The Philosophy of Adaptive Designs........................................... 238
C. Two-Stage Adaptive Designs ....................................................... 238
1. Adaptation ............................................................................... 239
2. Hypothesis Testing.................................................................. 240
3. Point Estimation and Confidence Intervals ............................ 241
D. Adaptive Group Sequential Designs ............................................ 242
V. Discussion ............................................................................................ 242
References........................................................................................................ 243
Some bias is inevitable in controlled clinical trials. The issue relevant to trial
modifications based on external or internal data is whether additional bias
would be introduced. It is naive to equate blindness with validity, i.e., to accept
the premise that the trial organizers are free to make trial modifications without
introducing bias if only blinded data are used, while bias is unavoidable but
immeasurable when modifications are based on unblinded data. We point out that
neither part of this proposition is true. A fruitful approach presented here is to
examine how trial modifications would introduce additional bias from various
sources. While careful planning and execution are necessary, it is also essential to
select appropriate adaptive design and inferential procedures in order to avoid
additional bias.