Design-Conditional Prior Elicitation for Dirichlet Process Mixtures: A Unified Framework for Cluster Counts and Weight Control

TL;DR

This paper addresses how to specify priors for the Dirichlet process concentration parameter in DP mixtures when the design size $J$ is fixed. It introduces Design-Conditional Elicitation (DCE), a workflow that translates beliefs about the expected number of occupied clusters $K_J$ (and its uncertainty) into Gamma hyperparameters for $\alpha$, using Two-Stage Moment Matching (TSMM) to achieve fast, accurate calibration. A key contribution is the Dual-Anchor diagnostic, which reports and optionally constrains the unintended weight concentration (dominance) that can arise even when $K_J$ is calibrated to a target. The framework is implemented in the DPprior R package, including reporting templates and reproducible diagnostics; simulations show that standard vague priors induce substantial cluster-collapse, while DCE with TSMM (and Dual-Anchor when needed) substantially improves cluster recovery and provides transparent trade-offs between partition fidelity and weight concentration. The work has practical impact for fixed-design educational and behavioral studies, multisite trials, and meta-analyses by enabling principled, interpretable prior specification and reproducible reporting of prior implications on both partition structure and mixture weights.

Abstract

Dirichlet process mixture (DPM) models are widely used for semiparametric Bayesian analysis in educational and behavioral research, yet specifying the concentration parameter remains a critical barrier. Default hyperpriors often impose strong, unintended assumptions about clustering, while existing calibration methods based on cluster counts suffer from computational inefficiency and fail to control the distribution of mixture weights. This article introduces Design-Conditional Elicitation (DCE), a unified framework that translates practitioner beliefs about cluster structure into coherent Gamma hyperpriors for a fixed design size J. DCE makes three contributions. First, it solves the computational bottleneck using Two-Stage Moment Matching (TSMM), which couples a closed-form approximation with an exact Newton refinement to calibrate hyperparameters without grid search. Second, addressing the "unintended prior" phenomenon, DCE incorporates a Dual-Anchor protocol to diagnose and optionally constrain the risk of weight dominance while transparently reporting the resulting trade-off against cluster-count fidelity. Third, the complete workflow is implemented in the open-source DPprior R package with reproducible diagnostics and a reporting checklist. Simulation studies demonstrate that common defaults such as Gamma(1, 1) induce posterior collapse rates exceeding 60% regardless of the true cluster structure, while DCE-calibrated priors substantially reduce bias and improve recovery across varying levels of data informativeness.

Figures (23)

• Figure 1: The unintended-prior phenomenon: A $K_J$-calibrated prior can imply substantial dominance risk under the weight view. Panel A: target (circles) vs. induced (bars) distribution of $K_{50}$. Panel B: marginal density of $w_1$ with dominance-tail probabilities. $J = 50$; $\alpha \sim \mathrm{Gamma}(1.60, 1.22)$.
• Figure 2: Design-conditional elicitation (DCE) workflow. Practitioners specify $J$, target $\mu_K = \mathbb{E}(K_J)$, and uncertainty $\sigma_K^2$. TSMM calibrates $(a, b)$; diagnostics report $p(K_J)$ and $\Pr(w_1 > t)$. Optional Dual-Anchor refines when dominance risk exceeds tolerance.
• Figure 3: Posterior expected cluster count $\mathbb{E}[K_J \mid \text{data}]$ by method and scenario. Rows: $K^* \in \{5, 10, 30\}$; columns: $I \in \{0.2, 0.5, 0.8\}$. Boxplots across 200 replications; dashed lines indicate true $K^*$.
• Figure 4: Probability of cluster collapse $\Pr(K_J = 1 \mid \text{data})$ across conditions. Bars: mean collapse probability; error bars: SD across 200 replications.
• Figure E.1: Dual-Anchor Pareto frontier ($J = 50$, $\mu_K = 5$). The $x$-axis displays $|\mathbb{E}[K_J] - \mu_K^*|$; the $y$-axis displays $\Pr(w_1 > 0.9)$. Each point represents a calibrated hyperprior at a particular $\lambda$ value (color scale). The red triangle marks the $K$-only TSMM solution ($\lambda = 1$); the horizontal dashed line indicates the constraint $\Pr(w_1 > 0.9) = 0.10$; the gray-shaded region satisfies this constraint.

Theorems & Definitions (62)

• Theorem A.1: Poisson--Binomial Representation
• proof
• Theorem A.2: Antoniak Distribution
• proof
• Proposition A.1: Conditional Moments
• proof
• Corollary A.1: Conditional Underdispersion
• proof
• Theorem A.3: Marginal PMF
• proof