2.2. Statistical analyses

Chi-square tests were run to test differences between groups (alcohol-dependent vs. controls) in gender distribution.

A series of mixed-effect models were run to examine group, gender and group-by-gender effects for age, education, monthly standard drinks, monthly cigarettes and brain volumes.

This technique statistically accommodates dependency between observations in a nested design (i.e., participants within sites) (Aarts et al., 2014). Site was treated as a random intercept to account for the systematic site-level variation in the dependent variables expected to occur from differences in scanners, protocols and assessments. The extent of variation explained by site-level differences was estimated as an intra-class correlation (ICC).

First, we examined the impact of factors including group (control, alcohol-dependent), gender (man, woman) and group-by-gender on the volume of a-priori ROIs as dependent variables, accounting for age, education years and intracranial volume (ICV). Separate models were run for each ROI. Significant group-by-gender effects were interrogated using pairwise post-hoc comparisons. In the text, we expressed the significant difference between two mean volumes as a percentage (%) difference (formula: [(predicted mean 1 - predicted mean 2 / (predicted mean 1 + predicted mean 2) / 2) *100]).

Second, in people with alcohol dependence, we explored whether gender or monthly standard drinks (model A), gender-by-standard drinks (model B) or monthly standard drinks separately in men and women (models C and D) predicted the volume of those ROIs that demonstrated significant group-by-gender effects, controlling for age, education years and ICV.

Both analyses were replicated on a sensitivity subsample where men and women with alcohol dependence were matched by monthly standard drinks (SI Appendix, Table S3).

Additional analyses including tobacco use (i.e., monthly cigarettes) as a covariate, were run in a subsample where this data was also available (465 alcohol dependent participants and 140 controls). Tobacco use (i.e., presence versus absence) did not significantly affect the results and was not included as a covariate in final models (SI Appendix, Table S4).

Alcohol use was positively skewed (skewness = 2.40) and was square-root transformed (skewness = 1.03) prior to statistical analyses (SI Appendix, Figure S1). Volumetric results were corrected for multiple comparisons using Benjamini and Yekutieli’s modified False Discovery Rate (FDR) method (Benjamini and Yekutieli, 2001, Newson, 2010) that was applied independently to each beta coefficient (for example the P-values for group comparisons were corrected separately from the P-values for sex comparisons, etc.). Cohen’s d was used to estimate effect sizes of the differences between groups, based on the marginal means predicted by the model. All statistical analyses were performed using STATA 14 (StataCorp; 2015).