Statistics

Statistics were performed using GraphPad Prism versions 7 and 8 (www.graphpad.com) with two-tailed hypothesis tests. All data were tested for homogeneity of variance and normality to choose the appropriate statistical test. Homogeneity of variance was assessed with an F test (two-group datasets) and a Brown-Forsythe test (> 2 group datasets) but not repeated measures; normality was always assessed with a Shapiro–Wilk test. Repeated measures data was tested for normality (sphericity was not assumed) with a Geisser-Greenhouse correction (see Supplementary Table ^S1 for analysis details). If the n was too low to test for normality (n ≤ 2 or identical values resulting in ≤ 2 unique n’s, i.e., in the serum data), data were analyzed non-parametrically. Log transformation was used to resolve heteroscedastic and non-normally distributed datasets, with the exception of the microvesicular and macrovesicular steatosis scores (due to zero values) and body weight/fat/lean change data (due to negative values). If the log-transformed data was non-normal, or if it was only heteroscedastic but the matching raw data was homoscedastic, the matching raw data was analyzed non-parametrically. If the log-transformed data was heteroscedastic but resolved the non-normal distribution of the raw data, the log transformed data was analyzed with the appropriate correction or test adjusting for heteroscedasticity. And if the log-transformed data was both homoscedastic and normally distributed, it was analyzed with the appropriate parametric test.

Data with two groups that were normally distributed and/or homoscedastic were analyzed with parametric (unpaired t-test) or non-parametric (Mann–Whitney) tests, and a Welch’s correction was used to adjust for heteroscedastic data. Data with three or more groups (not repeated measures) that were normally distributed and/or homoscedastic were analyzed with parametric (one-way ANOVA) or non-parametric (Kruskal–Wallis) tests, and a Welch’s ANOVA was used for heteroscedastic data. Analyses including multiple comparisons each underwent the two-stage linear step-up procedure of Benjamini, Krieger and Yekutieli to correct for multiple comparisons. Data with repeated measures missing any data points or non-normally distributed were analyzed using a mixed-effects model ANOVA whereas complete and normally distributed data were analyzed with a two-way ANOVA with repeated measures; significant pairwise comparisons for repeated measures data can be found in Supplementary Table ^S1. Pearson’s R correlations were run on normally distributed and homoscedastic data, while Spearman’s rank-order tests were run on non-normally distributed and/or heteroscedastic data (see Supplementary Table ^S3).

For the metabolomics data (Supplementary Table ^S5), each metabolite was analyzed individually, then the main effect p values were subjected to a Benjamini–Hochberg false-discovery rate correction (set to 5%), and multiple comparisons of significant metabolites underwent the two-stage linear step-up procedure of Benjamini, Krieger and Yekutieli-correction. In addition to main effects across all groups, the LD and WD groups were analyzed separately and, as with the main effects of all groups, results were subjected to a Benjamini–Hochberg false-discovery rate correction (set to 5%). In some instances, results demonstrated a significant corrected p value (i.e., q value) but not a significant uncorrected p value, a phenomenon sometimes called a “paradoxical result”³⁸. In the studies herein, this exclusively occurred in the repeated measures data and the metabolomics data; paradoxical comparisons are reported in Supplementary Tables ^S1 and ^S5 and noted in the figure legends.