To assess whether it is feasible to elicit patients' preferences for treatments and then to proceed with randomisation which may allocate those with preferences to their less preferred treatment; and to describe which prognostic variables were associated with such preferences within the context of a randomised trial of an exercise programme for back pain. The first 97 patients enrolled in a randomised controlled trial (RCT) for the treatment of back pain were asked about their preferences, health characteristics and other prognostic variables. Fifty-eight (60%) patients preferred to be allocated to the exercise programme whilst 38 (39%) were indifferent; one patient preferred conventional general practitioner (GP) management. No patient refused randomisation. Comparing patients preferring the exercise programme with indifferent patients showed that the former had a higher belief in the effectiveness of the new treatment (P < 0.01), tended to have worse back pain (P = 0.09), had back pain for a shorter duration (P = 0.04), and tended to have had more GP home visits (P = 0.06). For many randomised trials preference may be an important prognostic variable. In such circumstances, preference should be taken into account in the final analysis. This study demonstrates it is sometimes feasible to randomise patients to their less preferred treatment, thus allowing more robust statistical comparisons between randomised groups. This modification may make RCTs more rigorous and improve their external validity.