It is well established that random assignment between experimental treatment and control arms is the gold standard in clinical trials to minimise differences between the groups being compared and safeguard against bias. There is, however, a fear that such random allocation may not accord with patients’ preferences for the intervention or treatment, thereby compromising trial validity. It is possible that patients may resent not receiving their treatment of choice, and their negative attitude may lead to non-adherence to treatment or affect outcomes in some other way. Consequently, one option for trial designers is to include patient preference arms, whereby patients with no treatment preference are randomly allocated to experimental and control arms, while patients expressing a treatment preference are allocated to receive their preferred treatment.¹

Most of the debate about preference arms has focused on issues concerning increases in the sample size required, the types and stages of randomised design, and how to compare the groups. Ethical issues also arise, leading to concerns about those with strong treatment preferences being included in studies in the first place, especially those in which randomisation takes place irrespective of patients’ elicited preferences (that is, to measure whether those who received their preferred treatment, or not, differed from the no preference groups). There is less information about how doctors’ preferences for treatment influence their patients’ preferences.²

A recent systematic review of the impact of patient and doctor preference for intervention in randomised trials, across a range of disorders, reported that elicitation of preferences led to a substantial proportion of potential participants (patients, including parents, and doctors) refusing randomisation. However, differences in outcome across the trials between randomised and preference groups were reported to be small, and where preference effects were noted these were generally inconsistent.³ The authors of this study concluded that there was little evidence that preference substantially interfered with the internal or external validity of trials. The review included studies according to methodological criteria and if they “measured or recorded patient or physician preference” (page 1090). However, the reviewed studies varied widely in their approach to eliciting preferences, and preferences often appeared to have been simply noted in the consent to randomise procedures used. The methods used in most research on preferences include single item questions asking patients to state the option they would choose, Likert scales of the strength of the stated preference, and utility measurements (for example, rating scales, time trade offs, standard gamble methods)²; less commonly, preferences have been explored using discrete choice analysis.⁴ The question arises: does it really matter if preferences are not rigorously assessed in trials?

DO PREFERENCE ARMS REALLY REPRESENT PATIENTS’ PREFERENCES?

It matters if preferences are not rigorously assessed because of the review’s finding that there was little evidence that preference substantially interfered with trial validity. While this may be a valid conclusion, the result could also be simply attributable to each preference arm containing bias or a random mix of genuine preferences. Arguably, a preference can only be “true” if based on full, clear, and unbiased information about the treatment options: a preference is a value for alternative options for action after informed deliberation of their risks and benefits. And preferences need rigorous, standardised elicitation: research has shown that preferences are complex phenomena that may be informed by a wide range of constructs. That is, patients may have different reasons for expressing similar preferences—reasons based on past, or close others’ experiences; on misunderstandings of the different treatments; or on a plethora of other factors that range from justified fear to considered practicality.⁵ Patients can be affected differently by the same situation,⁶ and attach different significance to the same processes and outcomes,⁷ resulting in different treatment preferences. Additionally, when confronted with one or more unfamiliar treatments and asked to state a preference, a patient’s immediate answer might not survive further reflection, or indeed, first contact with the treatment itself. Furthermore, we must be cautious in interpreting preference studies even when these do provide patients with apparently full descriptions of treatment options. It is well known that patients’ and doctors’ preferences for treatment can be influenced by “framing” effects (the presentation of information in different ways, negatively or positively, and in a different order), understandings of the concept of risk (relative or absolute risks); verbal (“rarely”, “sometimes”, “never”) or numeric (single figures, decimals, fractions, or percentages) descriptions of risks and benefits.⁷ Other influences might include recall bias in relation to the information on risks and benefits received; varying expectations; preference for the status quo or familiar scenarios; and, in the case of patients, by their doctor’s preferences, perhaps because of patients’ deference to doctors’ dominance in the decision making process and the inclination of a substantial minority, especially older patients, to leave the decision “to the doctor”⁸ (despite the fact that doctors’ and patients’ treatment preferences, when measured in a standardised way, have been shown to vary).² All this can mean that the relation between stated and actual “true” preferences is weak. Surprisingly, there is little research examining the relation between patients’ anticipated preferences for treatment, actual behaviour (treatment), and post-treatment preferences.⁹ Given these potential influences, then, it is not unexpected that the conclusion of one systematic review was that, while provision of information on risk reduced decisional conflicts and stimulated patients to be more active in treatment decisions, the effect on the outcome of these decisions (actual treatment) was uncertain.¹⁰

In view of all this potential “noise” in preference elicitation, it might be considered as surprising if there was evidence that preferences, as currently measured, substantially interfered with trial validity. The possibility that preference arms in trials may not reflect true, informed, rigorously assessed preferences means that the results of analyses of the effects of preferences on outcome are ambiguous. Asking a person if they prefer “tea or coffee” may result in a valid choice, given that such preferences are likely to be genuinely informed by familiarity with, and experience of, the two options. But in the case of treatment preferences, initially low patient knowledge of the processes, risks and benefits of the treatment options, means that such simplistic methods of questioning, which do not even consider the reasons underlying preferences, may be inappropriate for this complex topic. The conclusion can only be that the results of analyses of preference arms in trials must be viewed with caution, and the elicitation of preferences deserves more respect, with the application of better methodology to match the rigour of other aspects of trial design. If each preference arm in clinical trials does simply contain a random mix of genuine preferences, then this not only has implications for the rigour of research methodology, but also for the soundness of health policy on which such research is based.