Correction

Luis Gruber, Gregor Kastner, Anirban Bhattacharya, Debdeep Pati, Natesh S. Pillai, David B. DunsonJournal of the American Statistical Association 2025 · DOI · arXiv

Abstract

Bhattacharya et al. (2015, Journal of the American Statistical Association 110(512): 1479-1490) introduce a novel prior, the Dirichlet-Laplace (DL) prior, and propose a Markov chain Monte Carlo (MCMC) method to simulate posterior draws under this prior in a conditionally Gaussian setting. The original algorithm samples from conditional distributions in the wrong order, i.e., it does not correctly sample from the joint posterior distribution of all latent variables. This note details the issue and provides two simple solutions: A correction to the original algorithm and a new algorithm based on an alternative, yet equivalent, formulation of the prior. This corrigendum does not affect the theoretical results in Bhattacharya et al. (2015).

Introduction

In high-dimensional Bayesian statistics, sparsity is a key assumption: among thousands of potential predictors, only a small subset is typically relevant. The Dirichlet-Laplace (DL) prior is a powerful shrinkage tool designed to concentrate mass near zero for noise while maintaining heavy tails for true signals. To implement this prior, researchers utilize Markov chain Monte Carlo (MCMC) algorithms. However, a critical flaw was discovered in the original implementation: the sampler failed to follow the "marginal first, conditional later" principle during its blocking and marginalization steps. Consequently, the original algorithm did not draw from the correct joint posterior distribution.

Methods

The authors propose two distinct remedies to ensure valid posterior inference:

  1. Corrected Original Algorithm: This approach maintains the original variable set but reorders the sampling sequence. By ensuring that global and local scale parameters ($\phi$ and $\tau$) are sampled prior to the auxiliary variables ($\psi$), the dependency structure of the joint posterior is correctly respected.
  2. Alternative Formulation: This method reformulates the DL prior using independent Gamma distributions ($\lambda_j \sim \mathcal{G}(a, 1/2)$). This version is mathematically equivalent to the original DL prior but yields a simpler, more robust MCMC sampler that avoids the ordering complexities of the first approach.

Validation was performed using the "Getting it right" test (Geweke, 2004), which compares MCMC output against a known ground truth via QQ-plots.

Results and Discussion

The impact of the sampling error is most pronounced in high-dimensional settings. For instance, in the $DL_{1/n}$ configuration with $n=100$, the original algorithm's squared error was 42.47, compared to 16.37 for the corrected version.

Figure 1: QQ-plots comparing the original (top), corrected (middle), and alternative (bottom) algorithms. The original algorithm shows a distinct "kink," while the proposed fixes follow the 45-degree line.
Figure 2: Comparison of estimation errors for different sample sizes (n) and sparsity levels (qn/n). The corrected and alternative algorithms consistently outperform the original implementation.

The visual evidence in Figure 1 clearly demonstrates the failure of the original sampler to match the target distribution, whereas both proposed corrections align perfectly with the theoretical quantiles. Furthermore, the alternative formulation appears particularly stable when the concentration parameter $a$ is small, likely due to better numerical mixing.

Conclusion

While the theoretical properties of the Dirichlet-Laplace prior remain valid, the original MCMC implementation was flawed. The corrections provided in this note are essential for practitioners to obtain accurate posterior estimates. Future research may further investigate the numerical advantages of the alternative formulation in more complex hierarchical models.

Summary updated

Abstract

Bhattacharya et al. (2015, Journal of the American Statistical Association 110(512): 1479-1490) introduce a novel prior, the Dirichlet-Laplace (DL) prior, and propose a Markov chain Monte Carlo (MCMC) method to simulate posterior draws under this prior in a conditionally Gaussian setting. The original algorithm samples from conditional distributions in the wrong order, i.e., it does not correctly sample from the joint posterior distribution of all latent variables. This note details the issue and provides two simple solutions: A correction to the original algorithm and a new algorithm based on an alternative, yet equivalent, formulation of the prior. This corrigendum does not affect the theoretical results in Bhattacharya et al. (2015).
BibTeX
@article{gruber2025correction,
  author = {Luis Gruber and Gregor Kastner and Anirban Bhattacharya and Debdeep Pati and Natesh S. Pillai and David B. Dunson},
  title = {{Correction}},
  journal = {Journal of the American Statistical Association},
  year = {2025},
  doi = {10.1080/01621459.2025.2540256},
}
Figure