Abstract
This paper addresses the computational challenges of estimating multivariate probit (MVP) models in panel data settings. We focus on sampling high-dimensional correlation matrices and improving the overall efficiency of data augmentation. By reparameterizing the correlation matrix and employing Hamiltonian Monte Carlo (HMC) sampling, alongside a novel deterministic antithetic variable method for random effects and regression coefficients, we achieve significant efficiency gains. The methodology is applied to stated preference data from Australian general practitioners (GPs) regarding contraceptive products, revealing substantial practice variation and resistance to discussing certain long-acting reversible contraceptives (LARCs).
Introduction
Multivariate Probit (MVP) models are essential for analyzing correlated binary outcomes, such as a clinician's decision to discuss multiple treatment options. However, their estimation is notoriously difficult due to two primary hurdles: the strict positive-definiteness and unit-diagonal constraints of the correlation matrix, and the slow mixing of traditional MCMC samplers in high-dimensional latent variable spaces. This work introduces a robust Bayesian framework designed to overcome these bottlenecks.
Methodology
Unconstrained Reparameterization for HMC
To navigate the constrained space of correlation matrices, the authors map them into an unconstrained Euclidean space using a Cholesky factor transformation. This allows the use of the No-U-Turn Sampler (NUTS), a variant of Hamiltonian Monte Carlo (HMC). Unlike standard random-walk Metropolis-Hastings, HMC utilizes gradient information to explore the posterior distribution more effectively.
Deterministic Antithetic Sampling
A key innovation of this paper is the introduction of a deterministic antithetic sampler for regression coefficients and random effects. Instead of a stochastic update, the authors use a reflection step: $\psi(\theta) = 2\mu_\theta - \theta$. This induced negative correlation between successive MCMC samples significantly reduces the variance of the estimators, leading to "super-efficient" mixing.
Results and Clinical Application
Computational Efficiency
The proposed method demonstrates a 16-fold improvement in MCMC mixing performance compared to standard independent sampling. Integrated Autocorrelation Time (IACT)—a measure of how many MCMC steps are needed to get one independent sample—was reduced by factors ranging from 3.3x to 4.8x for critical model parameters, allowing for faster and more reliable inference.
Insights into GP Decision Making
Applying this model to a dataset of 162 Australian GPs evaluating 2,592 patient vignettes revealed significant "medical practice variation." While the contraceptive pill remains a standard discussion point, there is wild variation in the propensity to discuss Long-Acting Reversible Contraceptives (LARCs) like implants or IUDs.
Key findings include:
- Product Bundling: High posterior correlation (0.59) between discussing hormonal injections and implants suggests GPs view these as a single category of options.
- Systemic Resistance: Variation in LARC discussion often persists regardless of clinical indications, suggesting a lack of familiarity or systemic resistance among some practitioners.
Conclusion and Limitations
While highly efficient, the antithetic sampler is sensitive to initial values, requiring a standard "burn-in" period before switching to the deterministic proposal. Additionally, the current implementation assumes conditional posterior normality. Future work will explore relaxing the exogeneity assumption of patient characteristics and extending the antithetic method to non-normal posterior approximations. This framework provides a powerful tool for researchers analyzing complex, correlated decision-making processes in panel data.
Summary updated