Method

Participants were randomised to intervention or control procedures, repeated four times over consecutive days. The intervention procedure used feedback designed to increase the perception of happiness over sadness in ambiguous facial expressions, whereas the control procedure was identical but used feedback designed not to modify this perception. Participants completed self-report measures of depressive symptoms and mood before and after the training procedure, and 2 weeks after the end of the intervention. The study was approved by the University of Bristol Faculty of Science Research Ethics Committee (trial registration: ISRCTN02532638).

Young adults (n = 193) aged between 18 and 40 years (primarily undergraduates) recruited from the general population by advertisement completed the Beck Depression Inventory II (BDI-II). ^{Reference Beck, Steer and Brown5} Of these, 80 (70% female) scored 14 or more (reflecting mood disturbance) and agreed to take part in the study. All provided informed consent. Testing took place in the School of Experimental Psychology at the University of Bristol.

A computer-generated randomisation sequence was minimised by BDI-II score (<20 or ⩾20) to balance participants with mild (14–19) and moderate to severe (20+) scores in each group, and concealed by requiring the researcher to complete a web form. This generated a number that the researcher entered into the computer delivering the intervention, which in turn assigned the participant to either the intervention or control condition. Participants and the researcher were masked to treatment allocation.

Prototypical ‘happy’ and ‘sad’ composite images were generated from 20 individual male faces showing a happy facial expression, and the same individuals showing a sad expression from the Karolinska Directed Emotional Faces, ^{Reference Lundqvist, Flykt and OUhman6} using established techniques. ^{Reference Tiddeman, Burt and Perrett7} These were used as end-points of a linear morph sequence, that changed in displayed emotion incrementally from unambiguously ‘happy’, through ambiguity, to unambiguously ‘sad’.

The BDI-II is a 21-item mood scale designed to measure depressive symptoms, with each item scored on a four-point scale. ^{Reference Beck, Steer and Brown5} The Positive and Negative Affect Schedule (PANAS) comprises two ten-item mood scales designed to measure positive and negative affect, with each item scored on a five-point scale. ^{Reference Watson, Clark and Tellegen8}

The intervention and control conditions consisted of three phases: baseline, training and test. This procedure was identical on all 4 days over which the training was delivered. In the baseline phase, participants were required to judge faces from a morphed sequence as either sad or happy (i.e. two-alternative forced choice). These data allowed us to calculate the balance point at which participants shifted from perceiving happiness to perceiving sadness in the presented face. This was then followed by the training phase that differed only in that feedback (i.e. ‘correct’, ‘incorrect’) was provided. In the intervention condition, feedback was provided based on a shifted balance point, so that participants were trained to judge expressions near the balance point that were previously judged as sad, as happy (online Fig. DS1). In the control condition, feedback was based on the same balance point calculated in the first phase, so that the judgement of ambiguous expressions would not change. In the test phase, the balance point was again calculated, using the same methods as in the baseline phase, to establish the success of the procedure in modifying the perception of ambiguous emotional expressions. Full technical details of the intervention (e.g. presentation times, experimental procedure) are provided in the online supplement.

Our primary outcome measure was BDI-II score at 2-week follow-up after completion of training. Secondary outcome measures were PANAS subscale scores ^{Reference Watson, Clark and Tellegen8} at 2-week follow-up and balance point after completion of treatment.

Linear regression was used to compare the intervention and control groups on primary (BDI-II) and secondary (PANAS, balance point) outcomes. Analyses were adjusted for baseline BDI-II score (i.e. the minimisation variable), and baseline score on the outcome measure for secondary outcomes. Analyses were conducted using PASW Statistics 18 on Windows. Our sample size provided 80% power, at an alpha level of 0.05, to detect a difference of six points on the BDI-II (assuming s.d. = 10) and five points on the PANAS (s.d. = 7).