Data synthesis
Meta-analyses were used to describe findings in the present review when data extracted could be used to calculate the standardized mean difference or risk ratios. We employed Comprehensive Meta-Analysis (CMA) software (version 4, professional, Biostat Inc. USA) to analyse the data. Random effect models were used in all the analyses. We used standardized mean difference (SMD) and their 95% confidence intervals (CIs) for continuous data. When the measurement only supplied dichotomous options, such as cognitive impairment or not, we used the Odds Ratio (OR) and their 95% CIs instead. Pooled random-effects 95% prediction intervals and heterogeneity statistics were calculated for meta-analysis where at least three studies were included. Heterogeneity was evaluated by I2, Tau2, and Q statistics and their p-values. When I2 > 75%, we conducted a series of subgroup meta-analyses by splitting data according to participants’ characteristics (such as sex and age at follow-up) or study characteristics (such as outcome measurement scale) to examine the source of heterogeneity. Forest plots were created to provide a graphical overview of the individual studies and syntheses.
To assess publication bias, we employed a random-effect model to generate funnel plots for meta-analyses. In addition, Duval and Tweedie’s trim and fill were used to estimate the number of missing studies that may exist and the effect that these studies might have had on their outcome.32
Two reviewers (X.Z. and M.S.) independently graded evidence according to the GRADE handbook33 . The strength of evidence was initially set as low and was rated up for 1) large effect sizes (Relative risk <0.5 or >2, and SMD<-0.25 or >0.25), 2) where a dose-response relationship was shown, and 3) effect of plausible residual confounding (such as parental education level )34 was considered. The strength of evidence was classified as very low, low, moderate, or high quality.