Bayesian nonparametric inference for panel count data with an informative observation process

Ye Liang, Yang Li, Bin ZhangBiometrical 2018 · DOI · arXiv

Abstract

In this paper, the panel count data analysis for recurrent events is considered. Such analysis is useful for studying tumor or infection recurrences in both clinical trial and observational studies. A bivariate Gaussian Cox process model is proposed to jointly model the observation process and the recurrent event process. Bayesian nonparametric inference is proposed for simultaneously estimating regression parameters, bivariate frailty effects, and baseline intensity functions. Inference is done through Markov chain Monte Carlo, with fully developed computational techniques. Predictive inference is also discussed under the Bayesian setting. The proposed method is shown to be efficient via simulation studies. A clinical trial dataset on skin cancer patients is analyzed to illustrate the proposed approach.

Introduction

In many longitudinal medical studies, researchers track recurrent events such as the regrowth of skin cancer lesions or the occurrence of infection bouts. However, researchers rarely have access to the exact timing of every event. Instead, they encounter "panel count data"—the total number of events accumulated between discrete, often irregular, check-up times.

A primary challenge in analyzing this data is the informative observation process. In an ideal setting, patient visits would occur at random intervals. In practice, however, a patient’s condition often dictates their visit frequency; those experiencing worsening symptoms may visit the doctor more often. If a statistical model assumes these visit times are independent of the disease progression, the resulting estimates will be biased. This paper introduces a robust Bayesian nonparametric framework designed to jointly model the event occurrences and the observation process, ensuring more accurate estimation and prediction.

Methodology

Bivariate Log-Gaussian Cox Process (LGCP)

The authors propose a joint modeling approach using a bivariate log-Gaussian Cox process. This framework defines two separate but correlated Poisson processes: one for the observation times (when the patient is seen) and one for the actual events (the recurrences).

To link these processes, the model utilizes a bivariate log-normal distribution for "frailties"—subject-specific random effects. This captures the inherent correlation where a "high-risk" patient likely experiences both a higher frequency of events and a higher frequency of clinical visits.

Nonparametric Baseline Intensity

Rather than relying on rigid mathematical functions or splines—which often require the manual and somewhat arbitrary placement of "knots"—the authors use Gaussian Processes (GP) to model the baseline intensity. This nonparametric approach allows the model to adapt flexibly to the functional form of the data, providing a smoother and more natural estimation of the underlying disease rhythm over time.

Computational Strategy: RMHMC

Sampling from high-dimensional Gaussian Processes is traditionally computationally expensive. To make this inference practical, the authors employ Riemann manifold Hamiltonian Monte Carlo (RMHMC). By utilizing the geometric curvature of the parameter space (the Fisher information matrix), RMHMC navigates the complex posterior distribution far more efficiently than standard MCMC methods, such as Elliptical Slice Sampling, leading to faster convergence.

Empirical Results

Simulation Studies

The framework was validated using three distinct simulation settings designed to test the model against misspecification. The results showed that the proposed method consistently outperformed existing frequentist and spline-based benchmarks (such as HSW and ZTS). Specifically, the model achieved lower Root Mean Squared Error (RMSE) and superior 95% coverage probabilities, particularly in scenarios where the observation process was highly dependent on the event process.

Skin Cancer Clinical Trial

The model was applied to a real-world dataset of 291 patients participating in a clinical trial for the drug DFMO, intended to reduce non-melanoma skin cancers. While the trial visits were planned for six-month intervals, actual visit times varied significantly.

Key findings from the analysis included:

  • Predictive Power: The initial number of tumors was identified as a highly significant predictor of future recurrences.
  • Treatment Effect: The DFMO treatment showed a marginal effect on reducing recurrences.
  • Pattern Recognition: The model successfully captured the periodic "spikes" in the six-month visit schedule while simultaneously accounting for the random noise of actual patient behavior.

Conclusion and Future Directions

This research demonstrates that joint modeling is essential when the observation process is informative. By combining Gaussian Processes for flexibility and RMHMC for computational efficiency, the authors provide a comprehensive tool for Bayesian inference in panel count data.

Future work may address the sensitivity of hyperparameters in the GP kernel, which remains a delicate aspect of the model. Additionally, the authors suggest extending this framework to multivariate panel counts, which would allow researchers to track multiple correlated symptoms—such as fever and cough in a vaccine trial—simultaneously.

Summary updated

Abstract

In this paper, the panel count data analysis for recurrent events is considered. Such analysis is useful for studying tumor or infection recurrences in both clinical trial and observational studies. A bivariate Gaussian Cox process model is proposed to jointly model the observation process and the recurrent event process. Bayesian nonparametric inference is proposed for simultaneously estimating regression parameters, bivariate frailty effects, and baseline intensity functions. Inference is done through Markov chain Monte Carlo, with fully developed computational techniques. Predictive inference is also discussed under the Bayesian setting. The proposed method is shown to be efficient via simulation studies. A clinical trial dataset on skin cancer patients is analyzed to illustrate the proposed approach.
BibTeX
@inproceedings{liang2018bayesian,
  author = {Ye Liang and Yang Li and Bin Zhang},
  title = {{Bayesian nonparametric inference for panel count data with an informative observation process}},
  booktitle = {Biometrical},
  year = {2018},
  doi = {10.1002/bimj.201700176},
}
Figure