Statistical Analysis
Meta-analyses using random-effects models were conducted for each outcome. Adjusted mean differences (MD) and odds ratios (OR) were pooled for continuous and categorical outcomes, respectively. Different scales of reporting diet quality across studies, including one absolute unit increase, one SD increase, tertiles, quartiles, quintiles were transformed to calculate the effect size in the top tertile of diet quality scores compared with the bottom tertile using methodology reported in previous studies19,30,31. There are two assumptions for this method: 1) diet quality scores are normally distributed and 2) the associations with the outcomes are log-linear. In a normal distribution, the means of the top and bottom tertiles, quartiles, quintiles of are 2.18, 2.54, 2.80 SD apart, respectively. log ORs and SEs were multiplied by 2.18/SD, 2.18, 2.18/2.54, 2.18/2.80 for the transformation from 1 unit, 1 SD, quartile, quntile, respectively, to estimates in tertile. Heterogeneity was assessed using the I2 statistic, Cochran’s Q test (P-heterogeneity) and by visual inspection of the forest plots. Sources of heterogeneity were assessed by conducting subgroup analyses when there was a substantial amount of heterogeneity (I2≥75% or P-heterogeneity<0.05).
Sensitivity analyses were performed by excluding one study at a time to evaluate the influence of any individual study on the pooled estimate. The robustness of the estimates was also examined with sensitivity analyses excluding moderate- and low-quality studies. Publication bias was assessed by visual inspection of a funnel plot and Egger’s test when at least 10 studies were included in the meta-analyses32.
Data analyses were performed using R version 4.0.3 (R Project for Statistical Computing). Two-tailed P values were used and P < .05 was considered statistically significant.