Authors: Zongyu Li, Mark Steel, and Zhiyong Zhang
Abstract
Traditional mediation models in both the frequentist and Bayesian frameworks typically assume normality of the error terms. Violations of this assumption can impair the estimation and hypothesis testing of the mediation effect in conventional approaches. This study addresses the non-normality issue by explicitly modelling skewed and heavy-tailed error terms within the Bayesian mediation framework. Building on the work of Fernández and Steel (1998), this study introduces a novel family of distributions, termed the Centred Two-Piece Student $t$ Distribution (CTPT). The new distribution incorporates a skewness parameter into the Student $t$ distribution and centres it to have a mean of zero, enabling flexible modelling of error terms in Bayesian regression and mediation analysis. A class of standard improper priors is employed, and conditions for the existence of the posterior distribution and posterior moments are established, while enabling inference on both skewness and tail parameters. Simulation studies are conducted to examine parameter recovery accuracy and statistical power in testing mediation effects. Compared to traditional Bayesian and frequentist methods, particularly bootstrap-based approaches, our method gives greater statistical power when correctly specified, while maintaining robustness against model misspecification. The application of the proposed approach is illustrated through real data analysis. Additionally, we have developed an R package FlexBayesMed to implement our methods in linear regression and mediation analysis.
Introduction
In disciplines ranging from psychology to epidemiology, mediation analysis is a cornerstone for understanding the mechanisms through which an independent variable influences an outcome. By identifying intermediate variables, or mediators, researchers can move beyond simple correlation to uncover causal pathways. However, the statistical validity of these models often hinges on the assumption that error terms follow a normal distribution.
In practice, real-world data frequently exhibits skewness (asymmetry) and heavy tails (outliers). When these characteristics are present, standard models—whether frequentist or Bayesian—often suffer from biased estimates and reduced statistical power. While non-parametric methods like bootstrapping offer some robustness, they lack the efficiency of a well-specified parametric model. This paper introduces a flexible Bayesian framework designed to handle these complexities directly.
The Centred Two-Piece Student $t$ Distribution (CTPT)
The core innovation of this research is the Centred Two-Piece Student $t$ Distribution (CTPT). This distribution is designed to be a "drop-in" replacement for the normal distribution in regression and mediation models.
1. Addressing Skewness and Tails
The model builds on the "two-piece" logic: taking a symmetric Student $t$ distribution and scaling the two sides of the mode differently. This allows for the explicit modeling of skewness. The Student $t$ base naturally accounts for heavy tails through its degrees-of-freedom parameter.
2. The Importance of Centering
A significant challenge with many skewed distributions is that their mean is not zero. In a regression context, if the error term has a non-zero mean, the intercept becomes confounded with the skewness and tail parameters, making results difficult to interpret and compare across studies. The authors mathematically shifted the distribution to ensure a zero mean, preserving the standard interpretation of regression coefficients.
3. Bayesian Inference and Stability
The authors utilize a Bayesian framework with standard improper priors. They provide rigorous mathematical proofs ensuring that the resulting posterior distributions are valid and stable, even when using these "vague" priors. This allows for robust inference on both the mediation effects and the distributional parameters (skewness and tail-heaviness).
Evaluation and Results
The CTPT framework was evaluated through extensive simulations and a real-world application.
Simulation Studies
The researchers tested the model against datasets with varying degrees of non-normality. The CTPT model consistently outperformed traditional Ordinary Least Squares (OLS) and bootstrap methods. For instance, in scenarios with small sample sizes ($n=50$) and high skewness, the CTPT model achieved a True Positive Rate (TPR) of 0.374, significantly higher than the 0.109 achieved by OLS-based bootstrap methods at the same false-positive rate. This demonstrates that correctly specifying the error distribution can dramatically increase the power to detect real mediation effects.
Real-World Application: NLSY79
The model was applied to data from the National Longitudinal Survey of Youth (NLSY79), examining how a mother’s education level influences a child’s mathematical achievement through the home environment. The CTPT analysis identified significant skewness in the data. It yielded a Bayes Factor of 13.021, providing "strong evidence" for a mediation effect that was less pronounced under standard normal assumptions.
Conclusion
The CTPT framework provides a powerful tool for researchers dealing with non-normal data in mediation analysis. By explicitly modeling skewness and heavy tails while maintaining a zero-mean error term, it offers both higher statistical power and better interpretability than traditional methods.
Future work may extend this framework to handle multimodal data or discrete/ordinal variables. To facilitate adoption, the authors have released the FlexBayesMed R package, making these advanced Bayesian techniques accessible to the broader research community.
Summary updated