I once read about a survey/polling methodology used by researchers
interested in truthful answers to particularly sensitive questions (it
may have been about date rape, but I am not certain of that).

  The approach involved having individual subjects generate a random
number known only to themselves ? e.g., with a die, roulette, or
equivalent device
type device. The subjects were instructed to either tell the truth or
just give a set answer based on the (secret) random number generated
(e.g., secretly cast a six sided die and tell the truth if the result
is <4. Otherwise just say "yes").

  Since researchers didn't know the specific random value generated by each
subject, the survey respondents could be assured of the
confidentiality of their individual answer. Since the researchers knew
the basic odds, however, they could answer population level questions
? e.g. what percentage of college students have committed date rape ?
with a much greater level of confidence then if they had just asked
the question and promised anonymity.

  Could someone providem me a pointer to a study that uses or defines
this methodology? Discussions of its effectiveness, caveats and design
tips would be particularly interesting.

Referred to as a "randomized response technique" or RRT.  Look up
randomized response technique in Google.

For example, see http://www.psychologie.uni-bonn.de/sozial/staff/musch/dis.pdf

The same site (from University of Bonn) has several articles on the
subject, including an approach that aims to deal with an anticipated
range of "cheaters" who still falsify answers in spite of the
confidentiality/protection that the RRT approach provides.

See ://www.google.ca/search?hl=en&lr=&ie=UTF-8&q=+site:www.psychologie.uni-bonn.de+randomized+response+technique