Data extraction and meta-analysis of data from all studies
Data were obtained from each study included in the systematic review and documented in contingency tables. We extracted the necessary data to calculate the incidence of PE in white women and in each other racial group. Whenever possible, we extracted the reported relative risk (RR) or odds ratio (OR) and 95% confidence intervals (CIs) from each study. Where available we extracted separate RR estimates with different degrees of confounder adjustment for the following prespecified conventional risk factors (age, weight and height or body mass index, smoking status and parity). First, we used raw data to adjust random effect models for meta-analyses using inverse variance method for pooling and DerSimonian-Laird to estimate the between-study variance (τ2). Second, we used adjusted OR from the included studies to also adjust the random effect model for meta-analysis with inverse variance for pooling but, in this case, we used restricted maximum-likelihood estimator (REML) for the between-study variance estimation. The pooled RR and/or pooled OR with 95% confidence intervals were estimated for race as a predictor for PE, using adjusted analysis as reported in the studies and a random effects model that considers both within- and between study variation30. Statistical heterogeneity among studies was evaluated using the I2, τ2 statistics and the p value of the Chi-Square test of Q31.
Publication bias, when the minimum number of included studies was 10, was assessed by plotting the RR estimate against precision (funnel plots) by Begg’s adjusted rank correlation test, and by Egger’s regression asymmetry test32,33.
Risk of bias assessment was made with quality in prognostic studies (QUIPS) tool34 presented and adjusted for this review. The following six domains were used: representativeness of study population, adequateness of follow-up period and attrition, appropriateness of racial origin classification, appropriateness of the definition of the outcome (PE), adequateness of statistical analysis and reporting. Each element was classified as low, moderate or high risk of bias. If two of the domains were assessed as having high risk of bias or four of the domains were assessed as having moderate risk of bias, then the overall risk of bias for a study was graded as high risk of bias. If three of the domains were assessed as having moderate risk of bias, or one domain was at high risk of bias and one was at moderate risk then the overall risk of bias was graded as moderate risk of bias. If all the domains within a study were graded as low risk of bias, or less than three were moderate and none was high, then the overall judgement for the study was low risk of bias.
Statistical software R 35 was used in all analyses, package ’meta’ 36 and ’metafor’ 37were used for the meta-analysis and package car 38 to clean the data.