Assessment and Definitions of Exposure, Mediator, Outcome, and Relevant Confounders

DNA samples from 1557 were analyzed; rs738409 was successfully genotyped in 100% of the samples. All samples were blindly duplicated to assess the reproducibility of the genotype, and reproducibility of 100% was obtained. The association between rs738409 and mediators/outcome was explored using an additive model (CC, CG, GG) of inheritance. A description of the rs738409 genotyping method is available in the Supplemental Material.

All liver biopsies were centrally read in a blinded fashion by the pathology committee of the NASH CRN as described previously. (27) NAFLD was defined as ≥5% of hepatocytes containing macrovesicular fat. Liver histology reports were based on the NASH CRN Scoring System, which assesses the severity of steatosis (0–3), lobular inflammation (0–2), hepatocyte ballooning degeneration (0–2), portal inflammation (0–2), and fibrosis (0–4). (27, 28) Fibrosis severity was considered the outcome of interest, whereas steatosis, lobular inflammation, hepatocyte ballooning degeneration, and portal inflammation were considered the mediators.

Before testing for mediation, potential correlations, and collinear effects between potential mediators (steatosis, lobular inflammation, hepatocyte ballooning degeneration, and portal inflammation) and fibrosis severity were checked using the multivariable linear regression models. The tolerance and the variance inflation factor (VIF), two collinearity diagnostic factors, were used to identify multicollinearity between mediators and the outcome. A small tolerance value, typically less than 0.1, might suggest strong multicollinearity. Although there is no formal VIF value for determining the presence of multicollinearity, VIF values that exceed 2.5 are often regarded as indicating multicollinearity. (29, 30)

Two different mediation models including either multiple mediators in parallel or sequentially were conducted to examine the effect of rs738409, by decomposing its total effect on fibrosis severity into direct and indirect effects, mediated by steatosis, lobular inflammation, ballooning, and portal inflammation. Regarding the analysis examining the effect of including mediators sequentially, we considered the inclusion of up to two consecutive histology traits (ie., steatosis → lobular inflammation, steatosis → ballooning, lobular inflammation → ballooning, etc.) The inclusion of 3 or more sequential traits did not improve model predictions (results are not shown). Overall, a mediation model displays 3 main types of path-specific estimates: 1) direct effect (represents the direct effect of rs738409 on fibrosis score, not operating through mediators “steatosis, lobular and portal inflammation and ballooning degeneration scores”), 2) indirect effect (represents the effect of rs738409 on fibrosis score that operates through mediators), 3) total effect (represents the sum of direct and indirect effects of rs738409 on fibrosis score). (31) Since age, gender, BMI, type 2 diabetes mellitus (T2DM), non-heavy alcohol intake, and other genetic variants (HSD17B13 rs72613567, TM6SF2 rs58542926, MBOAT7 rs641738) may be significantly associated with both the mediators and the outcome, all mediation analyses were adjusted by these relevant confounding factors.

Standardized regression coefficients express effect size measurements for total, direct, and indirect paths. Statistical significance for the total, natural direct, and natural indirect effects of rs738409 genotypes on fibrosis scores were determined using 95% bootstrap bias-corrected confidence intervals (CIs) based on 10,000 bootstrap samples. (32, 33) If the 95% bias-corrected CIs do not contain zero, the associations are considered significant. No missing observations were handled in our analyses. The mediation effects were estimated by linear regression-based mediation using the PROCESS macro, version 4.0, models 4 (parallel mediation) or 6 (sequential mediation), in SPSS (Version 27; Chicago, IL, USA). (33) The acyclic graphic or causal diagram (Suppl. Figure 1) shows our hypothesized associations between exposure, mediators, outcome, and relevant confounders. In the diagram, the indirect “mediation” effect is represented by the path from exposure to the outcome via the mediator (path a x path b). It expresses the fraction of the exposure effect that is mediated through a specific mediator. Mediation effects can be either positive or negative, depending on the signs of the a and b effects. When a is positive and b is negative, the mediation effect becomes negative. Path c represents the direct effect whereas the total effect is c’ (c + a x b). Additional details regarding our mediation models are given in the Supplemental Material.

Because our study population mostly comprised non-Hispanic (NH) White participants, the main analysis was performed in this ethnic group. However, to reproduce our results, a sensitivity analysis was conducted in a cohort including other races and ethnicities (n=404). A description of our selected cohort of NH-White has been previously published. (22, 23)

Monte Carlo power analyses as recommended by Schoemann et al. (34) with 10,000 repetitions, 20,000 Monte Carlo draws, and a 95% confidence interval, indicated 400 participants were needed to achieve between 0.79 and 0.84 statistical powers to detect a significant indirect effect in our model including 4 mediators in parallel. More details are given in Supplemental Material.

Data used in this study are deposited to the data coordinating center of the Nonalcoholic Steatohepatitis Clinical Research Network. Data can be made available by submitting a request to the NASH CRN data coordinating center, at https://jhuccs1.us/nash/ in compliance with its data-sharing policies and procedures.