Refurbishing the masked RCT design for psychological interventions

We would like to share some important statistical pitfalls of the randomised design in masked trials of music therapy such as that conducted by Erkkilä et al. ¹ The randomised controlled trial (RCT) is generally considered to be the optimal design for estimating treatment efficacy in medical interventions. In a double-blind RCT, the placebo effect is equally distributed between treatment groups. In Erkkilä et al’s trial, ¹ in the music therapy arm both the patient and the therapist became aware of the treatment that the patient was receiving well before total data had been collected. Thus, masking was jeopardised. Moreover, the authors did not allow for the patients’ treatment preferences. Patients who receive their preferred treatment may experience greater improvements in the outcome because of added motivation to follow the treatment protocol than patients who do not receive their preferred treatment.

Alternatives to the RCT design could have been used in the study. One option is the randomised consent design. In this, participants are randomised to treatment groups before the informed consent stage, and informed consent is then sought only for those allocated to the experimental treatment. ² Any sense of deprivation is less in the treatment as usual (TAU) group, as its members are unaware that they might have received a new treatment.

A second option is the partially randomised preference trial, in which participants without a treatment preference are randomised and those with a treatment preference are allocated to the treatment of their choice. This design has recently been used in some studies of psychological interventions for depression. The design has been recommended as it may improve both the internal and the external validity of clinical trials. ³ However, it may subject to the biases of an observational study and may not provide an unbiased measure of treatment effect. To improve both internal and external validity, Erkkilä et al’s RCT could have included a measure of preferences and detailed characteristics of those who refused to take part in the study because of the random allocation to treatment. This would have allowed the authors to measure preference effects at the analysis stage and to estimate the external validity of the trial.

A third option addresses the higher drop-out rate in the control group (11 v. 4) of the trial, which suggests the probably more demanding and careful follow-up in the experimental (music therapy) group. Here, instrumental variable methods have the advantage of allowing adjustment for non-adherence and loss to follow-up. Instrumental variables are associated with treatment choice (e.g. proximity to the music therapy clinic) but not with outcome. Had the patients’ treatment preferences been taken into account in this study, at least some of the eligible individuals would have refused to participate, especially those who lived further from the clinic. Instrumental variables provide an estimate of treatment effect that is adjusted for some of the bias associated with the patient preference design. ⁴

Last, it is worth mentioning the doubly randomised preference trial. ⁵ This is the most recently proposed method of estimating causal and preference effects. Patients are initially randomised to a randomisation arm, in which treatments are randomised, or to a preference arm, in which patients choose which treatment they receive.

These alternatives to the RCT, which are particularly appropriate for studies in which participants express a treatment preference or masking is less easy, are not free from biases. Nevertheless, they can ameliorate the external and internal validity of trials.