This was a double blind randomized controlled trial with the aim to transfer the promising results of the Amir
et al.[
17] and Schmidt
et al.[
18] studies to the Internet. Results did not confirm our hypothesis that Internet-delivered attention bias modification would be superior to a placebo condition. However, there was a significant time effect, but that is probably best explained by spontaneous remission since both the treatment and the placebo group displayed similar gains. In addition, the within group effect sizes were small. This is the second randomized trial not being able to replicate the initial promising results. However, the first non-replication by Boettcher and coworkers was not identical to the original studies since different stimuli were used [
30]. Specifically, the two original studies used the emotional facial expressions from Matsumoto and Ekman [
46], while Boettcher and colleges used the NimStim faces [
47]. The procedure in the present study mirrored the Amir study exactly including the same faces, number and spacing of the trials, and the location and duration of the stimuli. One obvious difference is that participants in the present trial did the training in a home setting as opposed to a university clinic ([
17,
18]). When looking at the proportion of correctly identified probes there was no difference and the marginally higher reaction time, as compared to Amir
et al.[
17] is explained by how elimination of individual trials were defined. The present study only excluded 3 SD above the individual trial mean (apx 1600 ms), while Amir
et al.[
17] simply excluded values higher than 1200 ms. Moreover, the method of delivery (
i.e., internet based flash program)
vs. personal computer delivery was a difference between current study and previous research.
An explanation of the null findings could be that this study also included participants with non-generalized social phobia. However, the mean scores on LSAS-SR were almost identical (73.4 in this trial compared to 74.5 in the Amir trial). In addition, when running the analysis with the NGSP and GSP groups separated, no differences in the treatment effect emerged.
When the participants who received the real treatment where asked to predict condition the absolute majority (81%) thought they had been randomized to the placebo group. One would think that such a low level of positive expectation would influence the trial. However, that cannot explain the lack of effect since the Amir
et al.[
17] and Schmidt
et al.[
18] papers reported almost identical numbers (78% and 94% respectively). Hence, there is a problem with the rationale and believability (
cf.)[
48]. However, the results when delivered on site have been impressive. Possibly attention training at home is negatively influenced by the fact that the person is sitting calmly in the comfort of his or her home. Perhaps there is an interaction between the active attention training and the mild anxiety that participants with social anxiety probably feel when they come to a university clinic to do their training (
cf.)[
49]. Although we instructed the participants to schedule the training when there would not be disturbed, there is a possibility that short breaks or non-focus reduces the effectiveness. Hence, the control over the procedures and the setting before, during and after training was suboptimal as compared to a lab. There could of course also be cultural differences. For example, in the case of panic disorder applied relaxation seems to work well in Sweden, but maybe not in the United Kingdom and USA [
50,
51]. The bias measure used in this study can be criticized. It could be argued that a task, in which the probe sometimes appears in the position of the disgust face and sometimes in the position of the neutral face, would be more accurate. However, Koster and coworkers [
52] proposed that the probe detection task may be modified such that vigilance for threat and disengagement from threat may be assessed by including baseline trials,
i.e., trial with two neutral faces. Using this new measure of bias, Koster
et al., found that individuals with anxiety have had difficulty in disengaging their attention from highly threatening pictures. This measure of bias has been used but other investigators to assess the specific componets of attentional bias in anxiety [
53‐
55]. The largest limitation is that participants were included without checking if they indeed had an initial attentional bias [
1]. This is a serious flaw since not all patients with presenting with anxiety show bias [
5]. However, the presence of a bias was not an inclusion criterion in the Amir
et al.[
17] or the Schmidt
et al.[
18] studies. Since there were no interaction effects in terms of change in bias between the two groups it can be concluded that not only was the internet-based treatment in the present study unable to modify attentional bias. In fact, there was the unexpected finding that larger bias was correlated with higher improvement. However, it should be noted that this was only a weak association, and could very well be explained as a random replicable finding. Indeed, as Heeren and coworkers have shown, engagement towards non-threat faces does not account for the positive treatment effects [
56]. It should be noted that the intervention employed only trains attention away from negativity. There could be added benefits of instead of training what is essentially avoidance, to also or instead train active selection of positive mood-supporting information [
57]. Perhaps the placebo condition, with equal frequency in the position of the threatening and neutral faces, is all that is needed to accomplish this disengagement from threat. Hence, future studies could, if done in an ethical way, test to include a condition of negative training.