---

THEORY

Semi-parametric statistics

Survival analysis

​Machine learning

Optimization/High-dimensional statistics

---

TOPICS

Causal inference

Data Integration

Personalized medicine

Interpretable variable selection

---

DATA

(Multi-center) clinical trials

Electronic health records (hospitals/claims/admin)

​The combination of above

​Longitudinal

---

DISEASE

Heart

Pediatric

Cancer

​Mental, etc.

---

Papers and softwares

Google scholar

Awards/Grants

Collaborators

Topics

GitHub

---

Glossary:

• ATE: average treatment effect
• HTE/CATE: heterogeneous (conditional average) treatment effect
• Effect modifier: the factor can modify the treatment effect in different population
• Common transportability condition: the outcome is independent of population source conditional on covariates. I.e., after adjusting for covariate, the outcomes in different populations have the same distribution.
• Non-parametric models: machine learning models that put no restriction on the data generating mechanism
• Green text: links to papers

1. DATA INTEGRATION
1.1 Robust trial augmentation: trials have limited efficiency due to their small sample sizes. We solve the challenge by integrating external data. 

• Study design and identification for various scenarios under common transportability conditions
• Hybrid control in early-phase trials

• Methods that using parametric, non-parametric, and foundation models that can ensure (asymptotically) unbiased estimation of the ATE of the trial, even if the common transportability condition fails and/or the models are misspecified

​1.2 Transportability: conclusions drawn from a trial are applicable only to its participants, while decision-makers often seek evidence relevant to a broader or specific population.

• ​A scoping review
• A method for transportability that assumes a more realistic condition---relative effect measures (w.g., risk ratio) are transportable​

​
1.3 Meta-analysis

• Causally-interpretable (network) meta-analysis to estimate the ATE, add-on effects, HTE, and subgroup effects (with R package and paper) when patients should take a regime from multiple treatments and patients in different populations have different access​ to the treatments

---

2. PERSONALIZED MEDICINE

• Effect score analyses: a method to estimate HTE of a trial, identify exceptional responders and optimal treatment strategies. It has valid statistical inference and is highly interpretable and clinically applicable even if with the use of any generic machine learning methods.
• HTE in longitudinal high-dimensional data
• Estimating subgroup and HTE in meta analysis

Figures by Wang et al. (JAMA) and Clemson University Open Textbooks

3. VARIABLE SELECTION

• Prior knowledge integration: incorporating clinician's knowledge to construct coherent and interpretable prediction models and identify clinically meaningful predictors. General framework; guide of using the overlapping group Lasso to achieve the goal for non-survival outcomes; for survival outcomes; R package.
• Confounder and effect modifier selection
• Applications

Figure by ​spot intelligence

Arriving at an interpretable predictive mechanism and a coherent prediction model with significant predictors.