Two investigators independently searched MEDLINE (January 1, 1946 to May 30, 2015), Embase (January 1, 1947 to May 30, 2015) and the Cochrane Central Register of Controlled Trials (CENTRAL; January 1, 1966 to May 30, 2015) without language or publication year restriction.

The following keywords, MeSH and free text were searched through MEDLINE: NAFLD, statin, and ezetimibe (Supplementary Table 1). Bibliographies of potentially relevant articles were manually reviewed to identify additional relevant studies. The identified articles were assessed individually for inclusion (Supplementary Table 2).

The studies were initially abstracted if they included the following keywords: NAFLD, statin, cholesterol lowering agent, ezetimibe. For inclusion, the studies were independently selected by two stages of screening using the Population Intervention Comparison Outcome framework [10]. Since the study objective was the histological effect with the lipid lowering agents, only those studies with liver biopsy results for diagnosis of NAFLD and post-treatment were included [11,12]. The required intervention included HMG-CoA reductase inhibitors or ezetimibe which can be administered at any dose for at least 6 months. The control group received no lipid lowering intervention or placebo, and there were no change of weight in all studies. The primary endpoint was improvement in hepatic steatosis while the secondary endpoint was improvement of NAS and safety.

Using a pre-defined data extraction form, two reviewers (H.Y.L. and D.W.J.) independently extracted data from each study. Any disagreement was independently reviewed by a third reviewer (H.J.K.). The following variables were extracted from the selected studies: (1) hepatic steatosis as evaluated by liver biopsy and/or quantitative fat measurement by magnetic resonance imaging (MRI); (2) NAS measurement before and after therapeutic intervention. All outcomes were assessed by differences between treated and control groups. The results were expressed as mean and standard deviations.

Two reviewers (H.Y.J. and D.W.J.) independently assessed the methodological qualities of included studies. The study quality was evaluated using the risk of bias by Cochrane for RCTs (Supplementary Fig. 1) and Newcastle Ottawa scale for NRSs (Supplementary Table 2) [13]. Any unresolved disagreements between reviewers were resolved by the third author (H.J.K.). Publication bias was not assessable due to the small numbers of studies.

We analyzed continuous data using standardized mean difference to combine trials that measure the same outcome but utilized different methods. The primary outcome was change of hepatic steatosis by liver biopsy (and MRI quantification in one study). The histological grading in NAFLD, inflammation, and fibrosis was based on scoring systems by either Brunt et al. [14] or Kleiner [15]. Secondary outcome were changes in NAS.

To assess for heterogeneity, we estimated the proportion of between-study inconsistency due to true differences between studies (rather than differences due to random error or chance) using the I2 statistic, with values of 25%, 50%, and 75% considered low, moderate, and high, respectively. Outcomes were analyzed using random effects model and standardized mean difference (SMD) to assess changes in measurements made by different scales. All analyses were performed using RevMan version 5.2 (http://community.cochrane.org/). This meta-analysis was conducted according to the Preferred Reporting Items for Systematic Revise and Meta-Analyses (PRISMA) statement.

Fig. 1 shows the details of literature research and selection process of the meta-analysis. The initial search strategy identified 857 articles. Of these, 667 publications were excluded after reviewing the title and abstract which indicated that they did not fulfill the selection criteria (Supplementary Table 3). For the remaining 39 articles [4,6-11,13-44], we performed full manuscript review and identified five relevant studies (three RCTs and two NRSs) to include in this meta-analysis.

Table 1 describes characteristics of the five included studies. The five studies comprised a total of 199 participants who received either statins (n = 47) or ezetimibe (n = 42) in the treatment group, and placebo (n = 97) or ursodeoxycholic acid (UDCA) (n = 13) in the control group. Ezetimibe was administered for 6 months and statins was administered from 6 months to 6 years [8,9]. Two studies were conducted in the USA, and one each in Japan, Sweden, and Romania. The study by Nelson et al. [4], Ekstedt et al. [28] and Georgescu and Georgescu [30] used statins, while the study by Loomba et al. [8] and Takeshita et al. [9] used ezetimibe as the lipid lowering agent. By inclusion criteria, liver biopsy was performed in all five studies, but magnetic resonance spectroscopy was also used in one study [8]. The control subjects received placebo except for those in the study by Georgescu and Georgescu [30] which used UDCA. In terms of race/ethnicity, three studies [4,28,30] included mostly or all Caucasian, while two studies [8,9] included mostly or all Asians. In both studies that used ezetimibe for 6 months, the baseline cholesterol level was within normal range (180 to 190 mg/dL). However, in the study by Georgescu and Georgescu [30] in which statin was used, the baseline cholesterol levels were high with mean level ranging 318 to 326 mg/dL. Loomba et al. [8] studied Caucasian subjects (baseline BMI, 33 to 34 kg/m²), while Takeshita et al. [9] studied East Asian subjects with lower BMI (baseline BMI, 28 to 31 kg/m²).

Among the three RCTs, the quality of two studies [8,9] was satisfactory, but one study [4] did not have random allocation sequence, optimal allocation concealment, and detailed data description. However, treated and control groups in all three RCTs were well-matched based on baseline characteristics with well-defined treatment response. In the two NRSs, subjects in the two groups were not well-matched since they were not randomized. The level of evidence and grade of recommendation for each outcome are summarized in Supplementary Table 2.

Cholesterol lowering agents did not significantly decreased the hepatic steatosis in NAFLD patients in the three RCTs (SMD, –0.10; 95% confidence interval [CI], –0.53 to 0.32) or in two NRSs (SMD, –3.00; 95% CI, –6.90 to 0.91) (Fig. 2).

Only two RCTs and one NRS reported the NAS data. Meta-analyzed result of the two RCT studies demonstrated a significant improvement of NAS (SMD, –0.30; 95% CI, –0.57 to –0.03). As shown in Fig. 3, pooled estimate of all three studies with available data also showed significant improvement. However, the mean reduction of NAS was modest: –1.0 in in Loomba et al. [8] and –0.65 in Takeshita et al. [9] but slightly higher in the case control study by Georgescu and Georgescu (mean, –1.7) [30].

There was no significant change in serum fasting glucose levels in the two RCTs (SMD, 0.20; 95% CI, –0.57 to 0.97) (Fig. 4) [8,9]. Data on glycated hemoglobin (HbA1c) changes were also only reported in these two RCTs and there was no significant change (SMD, 0.31; 95% CI, –0.15 to 0.76) (Fig. 5) [8,9].