Statistical Analysis
To investigate the causal relationship between obesity, mediators, and asthma outcomes, we conducted observational, MR two-stage least squares (2SLS) and MR sensitivity analyses for all relationships (Figure 1). Observational associations were analyzed using logistic regression models for binary asthma outcome/early puberty, and linear regression models were used for continuous outcomes, such as pulmonary function, FeNO, physical fitness, and SDB scores. The MR 2SLS analysis is a well-known method in MR for binary exposure and binary outcomes.29 The first stage involves the regression of the BMI/mediators on genetic scores, generating predicted values of BMI/mediators. The second stage involves a regression analysis of asthma on the predicted BMI/mediators. To test for directional pleiotropy, we used sensitivity analysis methods (MR inverse-variance weighted [IVW], Egger [slope], and weighted median method) to support the validity of causal inference from the MR analysis with multiple genetic variants.30 Observational mediation analysis was implemented using a mediation package for one mediator in R software.31 MR mediation analysis and proportion were calculated according to equations established by Burgess et al.14 The proportion of mediation was calculated as the estimate of the indirect effect divided by the estimate of the total effect, considering one mediator at a time. All model regression and sensitivity analyses were performed using R software (vers. 3.3.2, RStudio, Inc., Boston, MA, USA).
To evaluate the longitudinal effect of obesity on mediators, and of mediators on asthma throughout the three surveys in cohort 2, we used generalized estimating equations (GEE),32 which account for correlations between different measurements at different times in the same individual. A specific type of GEE with a Time-lag Model was used to approve temporal causality.33 In the Time-lag Model, predictors (adiposity status in the 2010 and 2011 surveys) were modeled using data that preceded the outcome variables (mediators assessed in 2011 and 2012); the model comprised the specific time-varying nature of adiposity status. The Time-lag Model also accounts for the temporal sequence in a possible cause-and-effect relationship.