Synthesis of results and summary measures
Extracted data from eligible studies were entered into
Revman™ v5.1 for data synthesis. For data presented as
median and interquartile range, estimates of mean and
standard deviation were determined using the method
described by Hozo et al. [7], which would reduce the
precision of our estimate of effect. Adjusted sample sizes
were calculated for cluster randomized trials by taking
into account the average cluster size and inter-cluster
correlation coefficients. Statistical heterogeneity was
assessed for each outcome of interest, and reported
using I
2
, with values greater than 50 % indicating
substantial heterogeneity. For outcomes not found to
have significant heterogeneity, summarized outcomes
(standardized mean difference for continuous variables
or relative risk for dichotomous variables) and 95 % confidence
intervals (CIs) were calculated using a randomeffects
model. For all tests, p values less 0.05 were
considered to be statistically significant. Where possible,
we performed sensitivity analysis by analyzing data on the
principle of intention-to-treat (ITT) and per-protocol analysis.
For ITT analysis, we compared patients randomized