Correction to “Alcohol-specific inhibition training in patients with alcohol use disorder: A multi-centre, double-blind randomized clinical trial examining drinking outcome and working mechanisms”

Abstract

Stein, M, Soravia, LM, Tschuemperlin, RM, Batschelet, HM, Jaeger, J, Roesner, S, et al. Alcohol-specific inhibition training in patients with alcohol use disorder: A multi-centre, double-blind randomized clinical trial examining drinking outcome and working mechanisms. Addiction. 2023; 118(4): 646–657. https://doi.org/10.1111/add.16104

In the context of secondary analyses [1] of this study's data, information relevant to the statistics presented in the original publication [2] came to our attention.

First, the multiple imputations used in the original publication in the regression analysis were distorted and do not replicate with newer versions of R and of the mice package [3]. Recomputing the regression analyses using imputations generated with the newer versions did not replicate the effects reported in the original publication; while similar on a descriptive level [with estimates for improved alcohol-specific inhibition training (Alc-IT) being superior to standard Alc-IT and control], no indicators for a significant effect of improved (β = 8.06, SE = 5.49, P = 0.145, 95% CI = −2.84 to 19.00) or standard Alc-IT (β = −2.22, SE = 5.66, P = 0.695, CI = −13.5 to 9.05) were yielded. We, therefore, must correct our statement that a significant effect of improved Alc-IT can be observed with a linear regression based on multiple imputations.

Importantly, the hierarchical linear model (HLM) results, which are not based on the multiple imputations and are, therefore, not affected by these corrections, still yield a significant effect of improved Alc-IT, as described in the original publication.

In such a case, maximum likelihood methods, like the HLM analyses, which are presented in section 2.3 of the Supporting information, are more appropriate [7, 8, 10-12]. We regret not addressing these issues more thoroughly before the original publication, as this would have led us to stick more tenaciously to the HLMs. These HLMs—originally presented as the main analyses by us—were moved from the main text to the Supporting information on intervention during the review process with the aim to enhance comparability with earlier studies. However, given the reasons above, it seems more important to analyze the data with the most appropriate approach, which is represented by the HLMs.

Second, in the sensitivity analyses, an error in the condition labels occurred, leading to control and improved Alc-IT being compared against standard Alc-IT as a baseline. Correct labels in eTable 2 would, therefore, have been as follows:

eTable 2(a): Analyses of PDA at 3-month follow-up under a MCAR and MNAR assumption (comparing control and improved Alc-IT against standard Alc-IT)

This analysis, therefore, indicates that improved Alc-IT performed significantly better than standard Alc-IT. If the analyses are computed comparing improved Alc-IT and standard Alc-IT against baseline, the effect of improved Alc-IT is insignificant, as can be seen in eTable 2 (b) below:

eTable 2(b): Analyses of PDA at 3-month follow-up under a MCAR and MNAR assumption (comparing standard and improved Alc-IT against control)

We, therefore, must correct our original statement that the sensitivity analyses also yielded significant differences between the groups receiving improved Alc-IT and control training. In the sensitivity analyses, these differences are only significant if improved Alc-IT is compared directly against standard Alc-IT.

As a consequence of these points, for which we apologize, we conclude that the effect of improved Alc-IT on percentage of days abstinent is less robust than stated in the original publication, because it only becomes statistically significant if the nested structure of the data is considered, as is the case in the HLMs. Given the statistical arguments listed above, which support our original plan to analyze the data with HLMs, we advise readers to consider the HLM results (that are listed in section 2.3 of the Supporting information), while of course acknowledging the fact that the insignificant effects of the regression based on the new imputations limit the robustness and generalizability of these effects.