Title: Latent multiple mediation analysis with the Bayesian lasso
Authors: Junhao Pan - Sun Yat-sen University (China) [presenting]
Lijin Zhang - Sun Yat-sen University (China)
Abstract: Mediators have played an essential role in helping researchers understand the mechanism through which the predictors affect the outcome variables. The existence of multiple mediators is also very common in behavior research. Traditional approaches for testing the indirect effects include the Sobel test, percentile bootstrap method, and bias-corrected bootstrap method. When handling multiple mediators simultaneously, the traditional approaches for testing the indirect effects, which include the Sobel test and bootstrap method, are prone to inflated Type I error rates and the overfitting problem. To provide a more effective variable selection tool in multiple mediation analysis, mediation models of observed variables were previously integrated with the frequentist Lasso (least absolute shrinkage and selection operator) method. However, this method has two limitations: (1) it does not take the measurement errors of manifest variables into account; (2) it cannot provide uncertainty information about indirect effects (e.g., interval estimation). The current study extended the Bayesian Lasso method into the framework of latent multiple mediation models to solve the above-mentioned problems. A Monte Carlo simulation study was conducted to compare the proposed method with the traditional Sobel and bootstrap methods. Recommendations and future directions were also provided based on the findings of the simulation study.