Statistical analysis
Meta-analysis was performed on crude data extracted from the text. We
calculated the weighted mean difference (WMD) for continuous outcomes
and the odds ratio (OR) for the dichotomous data, along with the 95%
confidence intervals (CIs). In instances in which a standard error was
not reported, we calculated the standard error of mean differences
according to the methods described in Cochrane Handbook[13].
Prior to analyzing the data, heterogeneity was assessed by the Cochran Q
test along with visual inspection of the forest plot. Then, it was
quantified by I2 test. A fixed-effect model was
used when the effects were assumed to be homogenous (P>0.05
or I2 <50 %)[13]. However, given
that the clinical settings differed across studies, we assumed the
presence of heterogeneity and used random-effects model in all
subsequent analyses, for the outcome of which were more conservative as
they consider differences both within and among studies in calculating
the error term used in the analysis[13].
Funnel plots were employed for detection of publication bias, in which
the effect sizes (e.g., OR or WMD) are plotted on the horizontal axis
and its variance (e.g., the standard error of the log effect) on the
vertical axis. Bias was revealed if the plots were asymmetrical about
the pooled value.
All statistical analyses were done with Review Manager 5.3.5 (Cochrane
Collaboration, Oxford, UK). Results were regarded as statistically
significant if P<0.05.