We conducted three versions of experiment 1 to test the replicability of our effect whilst adjusting for differences in task difficulty. Experiments 1a and 1b were identical, with the exception of the presentation speed which was slowed down from 83 (1a) to 100 ms (1b) in an attempt to equate task difficulty between the alcohol and non-alcoholic goals. Experiment 1c changed the non-alcohol stimulus category from pots/pans to shoes, for the same reason. Additionally, a larger sample was collected for experiment 1c in order to allow sensitivity to detect a potentially smaller stimulus-driven effect.
Methods
Participants
Table 1 presents participant’s characteristics. The inclusion criteria required that participants must have consumed alcohol in the last month, were not currently abstaining, and were from the University of Sussex student subject pool. These participants were remunerated with either partial course credit or small cash payment. Informed consent was collected prior to participation, and ethics were approved by the University of Sussex Ethics Committee in accordance with the 1964 Declaration of Helsinki.
Table 1 The mean demographic and questionnaire data from across all four experiments and standard deviations are presented in brackets
Experiments 1a and 1b were intended to test whether a goal-driven attentional bias to alcohol could be induced; therefore, sample size calculations were conducted prior to data collection using Gpower software to determine which sample size would be suitable to detect a goal-driven effect (Faul et al. 2009). This revealed that to detect an effect size of d = .92 (two-tailed; α = .05; 1 − β = .80), a sample of 12 participants was required. However 13 participants were originally recruited due to one being excluded due to a programming error. The expected effect size for this power analysis was taken from a previous demonstration of goal-driven attentional bias to emotional faces (Brown et al. under review).Footnote 1 The final sample size of Experiment 1b, after excluding one participant for currently abstaining from alcohol, was larger than 1a (n = 16) due to scheduling error.Footnote 2
The intention of experiment 1c was to test whether a stimulus-driven attentional bias was evident in the current paradigm. We, therefore, increased the sample size to detect a smaller alcohol bias effect which has been found in previous studies. A power analysis revealed that a sample of 60 participants should be suitable to detect a small alcohol bias effect of d = .37 (two-tailed; α = .05; 1 − β = .8). This effect size was based on the 95% lower bound confidence interval of the meta-analytically computed relationship between alcohol consumption and an “implicit” cognitive bias towards alcohol, as reported by Rooke et al. (2008).
Alcohol Use Questionnaire (AUQ)
The AUQ is a 12-item questionnaire which measures the frequency and speed of the weekly consumption of specific alcoholic drinks, which allows the computation of the number of units drank per week and binge score (Mehrabian and Russell 1978).
Alcohol Use Disorder Identification Test (AUDIT)
The AUDIT is a 10-item scale which measures not only both the frequency and amount of alcohol consumed but also the negative behavioural consequences from alcohol, e.g. when drinking is concerning to others (Saunders et al. 1993).
Anticipated Effects of Alcohol Scale (AEAS)
The AEAS is a 22-item scale that measures the expected emotions immediately after consuming an imagined amount of alcohol (four drinks for females and five drinks for males). The scale is composed of four subscales varying along dimensions of arousal and valence (Morean et al. 2012). The main subscale of interest was the positive high arousal factor, as this factor will indicate whether individuals perceived alcohol to be rewarding (cf. Bradley et al. 2001).
Stimuli
Across all experiments, stimuli were presented using E-prime 2.0 software on a Dell 1707FP computer. The resolution was set to 1280 × 1024, and the viewing distance was maintained at 59 cm using a chin-rest. Example stimuli are presented in Fig. 1, and all stimuli are available online via the Open Science Framework (OSF: osf.io/9n8yq). All target and distractor stimuli were images of single objects on a plain white background. The images within each category were selected so that they formed a heterogenous visual category with multiple features, textures, and shapes. The alcohol stimuli were selected so that there were equal numbers of exemplars of spirits, wine, and beers—and half of these stimuli were presented in glasses, the other half in bottles. Pots/pans images were selected, so that there were a variety of materials and colours which formed the category (e.g. ceramic, steel, and copper). Approximately half the exemplars were frying pans, the other half pots. The shoes were selected so that there were multiple different types of shoe (e.g. sports trainers, high heels, boots, and men’s formal shoes). Men’s shoes and women’s shoes were presented approximately equally, though there were some unisex shoes presented. These image selection criteria thus encouraged participants to form a search goal for a general category of objects, rather than any single feature.
The angles which the shoe and alcohol images appeared were more uniform than the pots/pans; we therefore rotated several exemplars from these categories, so that these categories were matched on the variability of stimulus orientation. The alcohol target category contained 12 full colour images of different types of alcohol. In experiments 1a and 1b, the non-alcohol target category contained 12 images of different types of pots/pans. In experiment 1c, the non-alcohol target category contained 12 images of shoes.
Three categories of distractor images were presented in each experiment: alcohol, pots/pans, and shoes. In experiments 1a and 1b, the shoe category was included as a completely goal-incongruent category (i.e. not matching either task search goal), whilst in experiment 1c the pots/pans were the goal-incongruent category. Each distractor category was composed of 16 images, which were visually similar to the target images of the same category but were never the same exemplars. All distractor and target images appeared an equal number of times within each condition. The distractors appeared to the left and right of the central stream with a gap of .5° between them. All centrally presented distractors measured 3.44° × 2.29°, whilst the parafoveal distractors measured 2.98° × 4.58°.
In total, 408 non-alcoholic filler images were selected to appear in the central stream. These were composed of 24 different everyday household objects with 17 different exemplars of each of these objects (see Appendix Table 3 for full list of non-alcoholic items stimuli). An additional 48 non-alcoholic object images were selected to appear as fillers in the parafoveal locations, these were composed of the same 24 object categories with two exemplars from each category. The parafoveal filler served to fill the other distractor location not occupied with an alcohol, shoe, or pot/pan distractor. All stimuli were sourced from Google images and appeared in isolation from other objects on a white background. During the task, these images were presented on a grey coloured screen (red/green/blue balance: 192, 192, 192). All images appeared four times across the experiment. Due to potential similarity to the shoe targets, in experiment 1c socks were removed from the filler stimuli and were replaced with 19 lamp images; 17 in the central set, two in the parafoveal set.
RSVP task
In experiment 1a, participants were instructed to search in a central RSVP stream of nine images for an object from a specific category, each image appeared for 83 ms. The task consisted of two blocks of 96 trials; in one block participants were instructed to search for “ALCOHOL”, in the other “POTS + PANS,” and this search order was counterbalanced between participants. Participants received 400-ms reminders of what the search goal was before each trial, i.e. “alcohol” or “pots and pans.” At the end of each trial, participants had to report whether they believed the target had been present or absent. Responses were made using the “c” and “m” keys, with the key-response assignment counterbalanced between participants. On half of the trials the target was present; the other half it was absent. The response screen contained only the words “present/absent?” and disappeared once the participants had responded.
When present, the target image could appear at positions five, six, seven, or eight in the RSVP stream. When absent that particular position in the stream was filled with a filler image. Distractor images appeared to the left and right of the central stream, one position was filled with either a shoe, pot/pan, or an alcohol distractor, whilst the other position was occupied with a filler image of the same size. Shoe, pot/pan, and alcohol distractors each appeared on a third of the trials in each block. These distractors always appeared two images prior to the target (i.e. lag 2). All within participants’ variables were counterbalanced within each block. Before the task started participants completed a 16-trial practice block of equal alcohol and pot/pans targets. Participants were verbally instructed before the main task that the target category would only vary between blocks, not between trials, and that the participants should ignore every image outside of the central stream.
Changes were made to experiment 1b due to the pot/pan targets being more difficult to detect than the alcohol targets in experiment 1a. We, therefore, slowed the stimulus presentation time down to 100 ms per image. This is more in line with previous RSVP tasks which have found implicit attentional capture by affective stimuli (Most et al. 2005). Despite the slower presentation time in experiment 1b, pot/pan targets were still detected less accurately than alcohol targets; therefore, we switched the non-alcoholic targets in experiment 1c to salient shoe images. The trials now started with an instruction to search for “SHOES” instead of “POTS + PANS.” The prompt in the response screen was also changed from “present/absent?” to a single “?” to avoid any influence of word order on responding.
Procedure
For experiments 1a and 1b, participants were tested in a dimly lit testing room at the University of Sussex. After providing informed consent, participants were given task instructions and then completed the practice block with supervision from the experimenter, after which they completed the RSVP task on their own. Participants then completed pen and paper versions of the AUDIT, AUQ, and AEAS in a random order. The experiment took approximately 25 min to complete. In experiment 1c, the procedure was identical to experiments 1a and 1b, with the exception that the questionnaires were presented using Inquisit 5 in order to automate randomisation of the questionnaire order. Half the participants completed the questionnaire prior to the RSVP task and half afterwards. Finally, participants were debriefed as to the full aims of the study.
Analytic strategy
Across experiments 1a, 1b, and 1c, we conducted the same analyses. The dependent variable used was A-prime (A′) detection sensitivity index which controls for response bias; this was computed based on the proportion of hits and false alarms made during the present/absent task response (Stanislaw and Todorov 1999; Zhang and Mueller 2005). A′ ranges from .5, which indicates that a signal cannot be distinguished from noise (i.e. chance detection), to 1, which corresponds to perfect detection of the target. In order to determine whether there was any significant difference in A′ across conditions, each individual study was initially analysed using a 2 × 3 repeated measures ANOVA in SPSS statistical software, using current goal type (alcohol/non-alcohol) and distractor type (alcohol/goal congruent non-alcohol/irrelevant non-alcohol) as the factors.
To follow up these comparisons and to determine the overall strength of the effect, we conducted pairwise comparisons across three studies using an internal meta-analysis. Four pairwise comparisons were computed; these were between the goal congruent distractors and the irrelevant distractor, in both search goal conditions (individual experiment comparisons are reported in Online Resource 1). The meta-analysis was conducted using the Metafor statistical package in R which weighted each experiment by its sample size (as described in Aloe and Becker 2012, Viechtbauer 2010). In all experiments, A′ scores were significantly skewed; therefore, a DerSimonian-Laird random effects model was used to compute the cumulative effects and confidence intervals, which is robust to violations of normality and is suitable for calculating cumulative effects from a small number of studies (DerSimonian and Laird 1986; Kontopantelis and Reeves 2012).
Bayes factors were calculated for all pairwise comparisons across experiments, as well as the cumulative effect. A Bayes factor compares evidence for the experimental hypothesis (positive attentional capture by alcohol versus an irrelevant distractor) and the null hypothesis (zero capture by alcohol versus an irrelevant distractor). The Bayes factor ranges from 0 to infinity. The strength of this evidence is indicated by the magnitude of the Bayes factor; values greater than three or less than .33 indicate substantial evidence for either the experimental or null hypothesis, respectively. A value closer to 1 suggests that the result is insensitive and any difference is “anecdotal” (Dienes 2008, 2011, 2014, 2016).
The Bayes factors were computed using a modified version of Baguley and Kaye’s (2010) R code (retrieved from Dienes 2008). To compute the factor, I used a half-normal distribution with a mean of zero to reflect the null hypothesis. The standard deviation of the distribution for all pairwise comparisons was set to .10, which is the plausible raw effect size for a difference between goal-congruent distractor and irrelevant distractor.Footnote 3 For meta-Bayes factors, used for the overall population mean, the effect was computed sequentially using Zoltan Dienes online calculator; first, combining the raw effect sizes and standard error of experiments 1a and 1b, then combining this cumulative posterior value with the mean and standard error of experiment 1c (Dienes 2008; Rouder and Morey 2011).