3.1 Dose-response meta-analysis
In order to analyze the correlation between BW and the risk of cancers, we used the dose-response meta-analysis to reflect the overall trend change of exposure (BW) level and the risk of outcome (cancer risk) indicators 46. Firstly, for all collected studies, we chose the most adjusted risk estimates and 95% confidence interval for the highest BW group versus the lowest group (reference). And the reported HRs and ORs were approximately considered RRs47. Then, we used both “Random-effects models” and “Fixed-effects models” to calculate the summarized RR estimates. If the heterogeneity was low (I2<50%), we used the value of fixed-effects models. Otherwise, we preferred to used random-effects models.
To estimate study‐specific dose‐response curves between BW and different types of cancer risk, we chose three models for fitting. The generalized least squares (GLS) model estimated the linear dose-response calculating the study‐specific RR of per 500g BW increment. The restricted cubic spline model was used to estimate the nonlinear trend of the dose-response relation 48. In the dose distribution, three knots were set to fit the model adjusting appropriately according to different cancer data. The accuracy of nonlinear fitting was assessed by the Wald test to determine whether the combined dose-response relationship is nonlinear. In addition, the quadratic model was also applied to estimate the nonlinear relationship between exposure and outcome which using the maximum likelihood estimation method as parameter estimation method. The heterogeneity across included studies was tested by the Q test and I2 test. We also tested the sensitivity by excluding one study at a time. The Egger’s test and the symmetry of the funnel plot were used to evaluate potential publication bias 49. All analyses were performed R (version 4.0.5) software with the packages of “dosresmeta”, “metaphor”, “mvmeta”, “rms” and “metafor”.