2.5. Statistical analysis

Data are presented as the mean (SD) or median (interquartile range, IQR) for continuous variables and percentages for categorical variables. Cross‐sectional group differences among FLI groups at baseline were assessed by unpaired t tests for normally distributed and loge transformed variables, by Mann–Whitney U tests for nonnormally distributed variables or by chi‐squared tests for categorical variables where appropriate. Multivariable linear regression analyses were carried out to disclose the associations of TMAO concentrations with clinical covariates and laboratory parameters, after adjustment for age and sex. To further evaluate whether TMAO was associated with FLI and HSI, two models were built including those variables associated with TMAO and mutually excluding FLI and HSI as well as its determinants. In order to identify the risk of multicollinearity in the multivariable regression analyses, the variance inflation factors (VIFs) were calculated. A high risk of multicollinearity was considered present if the calculated VIF was >5. ³¹

For the prospective analysis, the data of the two groups of participants: with NAFLD (n = 1598) and without NAFLD (n = 3694) were analysed separately, given the significant interaction between TMAO and NAFLD (P _int < .05). We plotted cumulative Kaplan–Meier curves for risk of all‐cause mortality during follow‐up according to tertiles of TMAO, in two groups of participants (with and without NAFLD). Time‐to‐event Cox proportional hazards models were used to compute hazard ratios (HRs) and 95% CI of all‐cause mortality risk in two groups of participants: with NAFLD (n = 1598) and without NAFLD (n = 3694). HRs were calculated in models adjusted for age, sex, T2D medication, smoking behaviour, alcohol consumption, history of cancer, systolic blood pressure, antidiabetic medication, lipid lowering medication, glucose, total cholesterol, high‐density lipoprotein cholesterol (HDL‐cholesterol), albuminuria, and eGFR at baseline. The Cox proportional hazard assumption was tested through the evaluation of independence between scaled Schoenfeld residuals with time for each variable and for every model as a whole; this assumption was met, with no indication for a violation. ³² In the two groups of participants (with and without NAFLD) the interactions of TMAO with age and eGFR were also evaluated. To further evaluate the robustness of the association and the risk of bias, a sensitivity analysis was conducted to calculate the Robustness of Inference to Replacement. ³³

All statistical analyses were performed with R language for statistical computing software, v. 4.0.3 (2020). ³⁴