In some randomized controlled trials, subjects with a better prognosis may
be diverted into the treatment group. This subverting of randomization resu
lts in an unobserved non-compliance with the originally intended treatment
assignment. Consequently, the estimate of treatment effect from these trial
s may be biased. This paper clarifies the determinants of the magnitude of
the bias and gives a sensitivity analysis that associates the amount that r
andomization is subverted and the resulting bias in treatment effect estima
tion. The methods are illustrated with a randomized controlled trial that e
valuates the efficacy of a culturally sensitive AIDS education video. Copyr
ight (C) 2001 John Wiley & Sons, Ltd.