Zian Zhuang
  • Home
  • Research
  • Publications
  • Teaching
  • CV

Research

Statistical methods for complex time-to-event data.

My main focus is prediction-powered and semi-supervised inference: using predictions from machine learning models and large language models to improve efficiency, while keeping inference valid under censoring and model misspecification.

Prediction-powered inference for survival outcomes

In many studies, the outcome is observed for only part of the sample, while predictions from machine learning models or large language models are available for everyone. With Dr. Gang Li, I am developing a semi-supervised Cox regression method that uses these predictions to gain efficiency without giving up valid inference.

The method targets a censoring-free, fixed-horizon Cox projection through inverse probability of censoring weighting, and it combines multiple, possibly dependent predictions with matrix weights. Under independent censoring, its asymptotic variance is no larger than that of the labeled-only estimator, whether or not the Cox model holds. Applications include NHANES mortality and dementia progression in data from the National Alzheimer’s Coordinating Center. A manuscript is in preparation, and a Python package is in development.

Next, I plan to extend this framework to causal estimands and to doubly robust handling of censoring, where the predictions also inform the model for the event time.

Evaluating risk prediction

To assess prediction models for right-censored competing-risk data, I proposed time-restricted generalized R² measures (Biometrics, in press). They give an interpretable measure of explained variation and model adequacy beyond the C-index, Brier score and time-dependent AUC.

I also wrote a tutorial on survival prediction performance measures (submitted to Statistics in Medicine) and maintain the R package PAmeasures. The package implements discrimination, calibration and overall accuracy measures for right-censored data, competing risks, and case-cohort and nested case-control designs. In earlier work, I contributed to an interpretable machine learning model for predicting antituberculosis drug-induced liver injury.

Causal inference with censored outcomes

I developed a data-adaptive censored two-stage least squares estimator for semiparametric accelerated failure time models (Statistics in Biosciences, 2026). It reduces endogeneity bias, improves efficiency, and runs more than 300 times faster than Bayesian parametric instrumental variable methods.

Clinical collaboration and epidemiological modeling

I collaborate with clinical researchers in the UCLA School of Medicine’s COPD Program, including analyses of the SPIROMICS cohort. Earlier, with Dr. Daihai He at The Hong Kong Polytechnic University, I used epidemic models to study COVID-19 transmission and estimate hospital bed shortages during the Wuhan lockdown, and I projected cancer mortality in Hong Kong.

Explore publications

© 2026 Zian Zhuang · Last updated November 2026