Corrigendum to “Causal relationships between substance use and insomnia” [Drug Alcohol Dependence 214 (2020) 108151] (Drug and Alcohol Dependence (2020) 214, (S0376871620303161), (10.1016/j.drugalcdep.2020.108151))

It has come to the attention of the authors that a correction is needed to our paper “Causal relationships between substance use and insomnia”. For our alcohol and smoking genetic instruments and outcomes we used summary statistics from the GSCAN meta-analyses (Liu et al., 2019). As the full summary statistics including data from 23andMe could not be provided by the GSCAN authors, we performed our own meta-analyses of 23andMe data (which we obtained from 23andMe through application) with summary statistics from all other GSCAN samples, excluding UK-Biobank. We could not replicate the exact method that GSCAN used for their meta-analyses (rareGWAMA), as that would have required information from the individual contributing cohorts. We rather used METAL to conduct the meta-analyses (Willer et al., 2010), and weighted the SNP-effects by the inverse of the standard errors. However, as the phenotype measurements in the original samples were heterogeneous, effect sizes and standard errors were not on the same scale, so SNP effects should have been weighted by their sample sizes. We re-conducted the meta-analyses, now based on sample size. As N-weighted meta-analysis results in z-scores rather than betas and standard errors, we had to derive these using procedures described before (Taylor et al., 2016). For beta, we used [Formula presented], with EAF representing the frequency of the effect allele in a European ancestry reference panel (The 1000 Genomes Project Consortium, 2015). The corresponding standard error was computed using [Formula presented]. We observed genomic inflation in the summary statistics for some of the traits, in which case the standard error was corrected using the LD-score regression intercept with [Formula presented] and corresponding p-values were computed. The genetic correlations between the old and new summary statistics were high (for smoking initiation rg = .98, cigarettes per day rg = .99, smoking cessation rg = .96, alcohol per week rg>.99). Using the new summary statistics, we recreated the genetic instruments and reran the MR analyses following identical procedures as before. The new instruments had slightly higher instrument R2 and showed similar genetic correlations with insomnia (see Table 1). Results from the MR analyses are presented in Table 2. Comparing the results with the original ones, there are some fluctuations in betas and p-values in both directions, but the overall patterns were the same; i.e., for all analyses the direction and significance of effects and remained the same, such that the new results do not lead to alternative interpretations. Additional statistics for robustness checks and sensitivity analyses can be found in the online updated Supplementary Material. The statistics show only minor differences from the previous results. One notable exception was the lower I2 for the analysis from smoking initiation on insomnia (indicating substantial NOME violation), so that MR Egger analysis for this relationship could no longer be interpreted. Figures S1a-S10a display the scatter plots for the MR analyses. Overall, we conclude that changes in results were negligible and did not merit different interpretations. Thus, the conclusions of our original publication stand: we found evidence for positive causal effects of liability to insomnia on all substance use phenotypes (smoking traits, alcohol dependence, cannabis initiation), except alcohol per week. Also, we found strong evidence that smoking initiation increased insomnia risk. The authors would like to apologise for any inconvenience caused.