2.4 Statistical analysis
The inverse variance-weighted (IVW) mode was applied as the primary approach, which used meta-analysis to combine the SNP-specific Wald ratio estimates for each IV and obtained an overall estimate of the effect17. Three other methods, MR-Egger18, weighted-median19, and weighted mode20approaches were used as supplements.
Heterogeneity was evaluated by Cochrane’s Q statistic21. The horizontal pleiotropy was accessed by the intercept of MR‐Egger regression, which should not be significantly different from 0 (i.e., P > 0.05)18. We also used the Mendelian randomization pleiotropy residual sum and outlier (MR-PRESSO) global test to identify and fix horizontal pleiotropic outliers22.
The multivariate IVW model with Lasso penalization was further used to evaluate the independent causal effect of related diseases on severe COVID-19. The Lasso penalization could shrink the coefficients of the invalid variables to zero, preventing overfitting23. In multivariate MR analysis, phenotypes with comparable traits could be treated as controls for each other to identify the predominant phenotypes24.
Odds ratios (OR) and 95% confidence intervals (CI) were used to describe the causal impact of exposures on outcomes. To correct for the multiple testing, the Bonferroni adjusted P values of significance were < 0.0167 (α = 0.05/3 exposure factors) and < 0.0125 (α = 0.05/4 exposure factors) in multivariate MR analyses for the three allergic diseases and the four asthma subtypes, respectively. P<0.0167/0.0125 was regarded as strong evidence of causality, while 0.0167/0.0125<P<0.05 was considered suggestive evidence of causality. In addition, other statistical tests had a two-sided design, with the threshold of statistical significance set at P<0.05. The R language (version 4.2.2) was used to select IVs, perform statistical analyses, and visualize our findings.