Statistical analysis

Data were analyzed using Stata version 12.0 (Stata Corporation, College Station, TX, USA). Before the data were synthesized, we first test the heterogeneity between the studies using Q chi-square test²⁴, in which a P value <0.10 was considered as significant heterogeneity. I² statistic was used to describe the percentage of the variability that attributed to heterogeneity across the studies rather than the chance. Studies with an I² statistic of <25%, ~50%, ~75%, ~100% are considered to have no, low, moderate, and high degree of heterogeneity, respectively²⁵. Pooled estimates were calculated using a fixed-effects model (Mantel–Haenszel method)²⁶; otherwise, a random-effects model (DerSimonian–Laird method)²⁷ was applied when significant heterogeneity among the included studies was found. If the heterogeneity was tested, subgroup analysis or sensitivity analysis was performed to explore the potential sources of heterogeneity.

Continuous variables, including mean changes from baseline in HbA1c, bodyweight, FPG, PPG, SBP, and DBP, were expressed as weight mean difference (WMD) with 95% confidence intervals (95%CIs); dichotomous variables, including the incidence of treatment-emerge adverse events, were expressed as relative risk (RR) with 95%CIs. The assessment of publication bias was evaluated by using Egger²⁸ and Begger²⁹ test. A P value less than 0.05 was judged as statistically significant, except where otherwise specified.