Study Population
In present study, we utilized data from six NHANES cycles (2005–2006, 2007–2008, 2009–2010, 2011–2012, 2013–2014, and 2015–2016) that involved a total of 60936 participants. We excluded participants younger than 20 years, leaving 34180 adults. Participants were excluded by the following criteria12,18–20: missing data for the CDAI calculation; missing data for the calculation of US fatty liver index (USFLI) score; missing data for alcohol consumption or presence of considerable alcohol consumption (> 21 drinks per week for male and > 14 drinks per week for female); participants with the hepatitis B surface antigen or hepatitis C antibodies; missing data for covariates. Therefore, a total of 9746 participants was included in the present study, and the flowchart of enrollment is presented in Fig. 1.
Definition of Primary Exposure
Diet affects the CDAI. CDAI data were derived from the two 24-hour dietary recall survey of NHANES participants. The first 24h was recorded face-to-face at a mobile examination center, and the second 24h was recorded by telephone 3–10 days later. The University of Texas Food Intake Analysis System and the United States Department of Agriculture Survey Nutrient Database were used to assess the dietary nutrient intake21. The nutritional estimates did not include any nutrients obtained from dietary supplements or medications. The average of the two 24-hour intakes was taken as the daily dietary intake for the present study22.
The CDAI was calculated from the mean dietary intake of vitamin A, vitamin C, vitamin E, zinc, selenium, and carotenoids obtained from two 24-hour recalls using a modified version developed by Wright et al.13. Six antioxidants were standardized by subtracting the mean and dividing by the standard deviation. Then the CDAI was calculated based on the sum of these standardized values.
$$CDAI= \sum _{i=1}^{6}\frac{{X}_{i}-{\mu }_{i}}{{s}_{i}}$$
In this formula, \({X}_{i}\) represents the individual daily intake of antioxidant components; \({\mu }_{i}\) represents the mean of \({X}_{i}\); \({s}_{i}\) represents the standard deviation for \({\mu }_{i}\)22–24\(\)
Definition of Outcome
NAFLD was defined according to the USFLI which was developed using the NHANES database, which was moderately improved accuracy compared to the Fatty Liver Index (FLI) in the multiethnic US population12. And it has been validated and used in several previous studies25–27. USFLI was developed based on race/ethnicity, age, gamma-glutamyl transferase (GGT), waist circumference (WC), fasting insulin, and fasting glucose with the following formula:
$$USFLI= \frac{{e}^{\begin{array}{c}(-0.8073\times non-Hispanic Black+0.3458\times Mexican American+0.0093\times Age\\ +0.6151\times \text{ln}GGT+0.0249\times WC+1.1792\times \text{ln}insulin\\ +0.8242\times \text{ln}glucose-14.7812)\end{array}}}{1+ {e}^{\begin{array}{c}(-0.8073\times non-Hispanic Black+0.3458\times Mexican American+0.0093\times Age\\ +0.6151\times \text{ln}GGT+0.0249\times WC+1.1792\times \text{ln}insulin\\ +0.8242\times \text{ln}glucose-14.7812)\end{array}}}\times 100$$
Scores range from 0 to 100. In this study, a USFLI score ≥ 30 was considered to have NAFLD as suggested by Ruhl and Everhart12, with an area under the receiver operating characteristic curve (AUROC) of 0.8 (sensitivity: 62%; specificity: 88%).
Covariates
Based on the literature, the following covariates were selected, including: age, sex, race/ethnicity, education level, marital status, family income, body mass index (BMI), alcohol drinking status, smoking status, diabetes, hypertension, metabolic syndrome, and total energy intake28,29. As used by NHANES, we divided race/ethnicity into Mexican American, other Hispanic, non-Hispanic White, non-Hispanic Black and Other (including multiracial). Education level was divided into two groups (high school or below and greater than high school). The marital status was classified as married, never married, living with a partner, and others (including divorced, widowed, and separated). Family income was categorized into 3 levels (< 1.3, 1.3–3.5, and ≥ 3.5) based on the family poverty income ratio (PIR). BMI was divided into 3 levels (< 25, 25–30, and ≥ 30 kg/m2). Alcohol drinking status was determined by the following survey question, “In any 1 year, have you had at least 12 drinks of any type of alcoholic beverage?” Participants who answered “yes” were defined as alcohol drinkers. Serum cotinine concentration was utilized as a proxy for environmental tobacco exposure and categorized into active/secondhand smoker (> 0.011 ng/mL) and nonsmoker (≤ 0.011 ng/mL). Diabetes was defined as using antidiabetic medication or a fasting glucose level equal or greater than 126 mg/dL. Hypertension was defined as using antihypertensive medication or average systolic blood pressure ≥ 140 mmHg and/or average diastolic blood pressure ≥ 80 mmHg. Metabolic syndrome was defined based on the Adult Treatment Panel Ⅲ criteria in 2005 as having at least 3 of the following: waist circumference greater than 102 cm in men or 88 cm in women, triglyceride level greater than 150 mg/dL, high-density lipoprotein cholesterol less than 40 mg/dL in men or less than 50 mg/dL in women, systolic blood pressure at least 130 mm Hg or diastolic blood pressure at least 85 mm Hg or taking hypertension medications, or fasting plasma glucose level at least 100 mg/dL or taking diabetes medications30,31. Total energy intake was calculated by averaging energy intake collected during two 24-h total nutrient recall interviews.
Statistical analysis
According to NHANES analytic guidelines, complex sampling design and sampling weights were considered in our analyses32. The characteristics of participants are described as means (95% CIs) for continuous variables and percentage frequencies (95% CIs) for categorical variables. Continuous data were compared using t tests, and categorical data were compared by the χ2 test. These means and frequencies can be generalized to the US adult population. No imputation method was used due to the percentage of missing data was small for any variable.
Odds ratios (ORs) and 95% CIs were calculated to assess the association between CDAI/antioxidant components and NAFLD using weighted logistic regression models. Given that the values of antioxidant components intake were skewed, a logarithmic change was performed before statistical analysis to ensure a normal distribution. And the CDAI was generally converted into categorical variables according to quartiles, and the P values for the trend were calculated. Three models were used in this study. Model 1 was the crude model with no covariates adjusted. Model 2 was adjusted for age, sex, and race/ethnicity. Model 3 was the fully adjusted model which including age, sex, race/ethnicity, education level, marital status, PIR, BMI, alcohol drinking status, serum cotinine, diabetes, hypertension, metabolic syndrome, and total energy intake.
In addition, interaction and subgroup analyses of association between CDAI and NAFLD were also performed according to sex (male, female), age (20–40, 40–60, ≥ 60 years), BMI (< 25, 25–30, and ≥ 30 kg/m2), diabetes (no, yes), hypertension (no, yes), and metabolic syndrome (no, yes) using logistic regression models.
To ensure the robustness of our research findings, we adopted the methods used by Ruan et al33. and conducted several sensitivity analyses. Initially, to assess the potential hepatotoxicity of certain pharmacological agents, we conducted a sensitivity analysis excluding participants who had been administered methotrexate, acitretin, pioglitazone, liraglutide, semaglutide, atorvastatin, or aspirin34–43. Subsequent to this, drawing upon extant literature which posited diminished prevalence rates of NAFLD and suboptimal diagnostic precision of USFLI within the non-Hispanic black cohort 12,44–46, we implemented a sensitivity analysis that omitted individuals belonging to this demographic. Finally, in an endeavor to mitigate the possibility of misclassification stemming from USFLI scores, we performed an additional sensitivity analysis utilizing the hepatic steatosis index (HSI)47. Here, we defined the NAFLD as an HSI score ≥ 36 in present study with the following formula:
$$HSI=8\times \frac{AST}{ALT}+BMI+2\left(if female\right)+2\left(if diabetes mellitus present\right)$$
Considering the calculation if HSI was based on BMI, BMI was not included as a covariate in this model to avoid the over adjustment. Ultimately, numerous previous studies have examined data from the NHANES to explore risk factors associated with various diseases, and some researchers have utilized weighted analysis methods, while others have employed unweighted approaches. Although NHANES employed complex sampling techniques to improve the representativeness and applicability of findings, conclusions derived from weighted and unweighted analyses can occasionally differ. Therefore, we conducted a sensitivity analysis using unweight data.
All statistical analysis was performed with R (version 4.1.3, R Project for Statistical Computing, Vienna, Austria) and EmpowerStats (version 4.1, Boston, Massachusetts). In all tests, P < 0.05 (2-sided) was considered to indicate statistical significance.