Albert and Chib (2001) Sequential Ordinal Modeling with Applications to Survival Data

Summary

Albert and Chib (2001) develop a Bayesian data-augmentation framework for the sequential ordinal model, in which an ordinal response $Y_i \in \{1,\ldots,J\}$ arises from a sequence of $J-1$ independent binary decisions — "exit at level $j$" vs. "continue to level $j+1$." This is equivalent to a discrete-time survival model and to the continuation-ratio logit of Agresti (1990). A two-block Gibbs sampler with truncated-normal augmentation handles both uncensored and right-censored observations, and marginal likelihoods computed via the Chib (1995) identity decisively favour the sequential model over cumulative probit and parametric survival alternatives in a hospital length-of-stay application ($N=1000$, $J=12$).

Key Claims

• The sequential model defines $J-1$ independent binary probit regressions; the discrete hazard is $\Pr(Y_i=j|Y_i\geq j)=\Phi(\gamma_j - \mathbf{x}_i'\boldsymbol{\delta})$.
• Cutpoints $\gamma_1,\ldots,\gamma_{J-1}$ need not be ordered (contrast with cumulative model); they are absorbed into an unrestricted coefficient vector $\boldsymbol{\beta}=(\gamma_1,\ldots,\gamma_{J-1},\boldsymbol{\delta}')'$.
• Data augmentation introduces $z_{ij}=w_{ij}-\gamma_j$ with $w_{ij}\sim\mathcal{N}(\mathbf{x}_{ij}'\boldsymbol{\beta},1)$; levels "passed through" contribute $\mathcal{TN}_{(0,\infty)}$ draws and the exit level contributes a $\mathcal{TN}_{(-\infty,0]}$ draw.
• Right-censored observations are handled by generating only the "passed through" draws with no exit draw.
• Marginal likelihood via Chib (1995): the Rao-Blackwellised posterior ordinate at $\boldsymbol{\beta}^*$ averages the Normal full conditional over the Gibbs draws of $\mathbf{z}$.
• Model comparison (6 competing models): sequential quadratic-baseline ($\ln m=-2094.0$) beats sequential basic ($-2098.6$), sequential interaction ($-2117.0$), cumulative probit ($-2146.7$), Weibull ($-2191.2$), and log-logistic ($-2223.2$).
• Best reduced model retains race (RACE), private insurance (PRIVINS), comorbidity index (COMORB), coronary artery bypass graft (CABG), percutaneous transluminal coronary angioplasty (PTCA) ($\ln m=-2092.1$); CABG has the largest effect ($\hat{\delta}=0.768$).

Concepts Introduced or Extended

• Sequential Ordinal Model
• Binary Probit
• Marginal Data Density
• Gibbs Sampler

Entities Mentioned

• James H. Albert
• Siddhartha Chib

Quotes

"The sequential model is of interest in survival analysis settings where the response is a discrete survival time." (p. 829)

"An important feature of this model is that the cutpoints $\gamma_1,\ldots,\gamma_{J-1}$ are not constrained to be in increasing order." (p. 830)

My Take

The paper is methodologically tight: the data-augmentation Gibbs from Albert-Chib (1993b) and the Chib (1995) marginal likelihood identity both transfer to the sequential setting with minimal modification. The hospital length-of-stay example makes a compelling practical case — the sequential model wins by Bayes factors on the order of $10^{20}$–$10^{22}$ over the cumulative and parametric alternatives, and the quadratic polynomial baseline improves parsimony further. The main limitation is the assumption of stage-invariant covariate effects; the interaction extension partially addresses this at the cost of multiplying parameters. The paper quietly unifies continuation-ratio logit, discrete-time hazard, and Bayesian ordinal modeling under one sampler.