About the Pan Group

The Pan Group develops statistical methods and computational tools to make clinical-trial evidence more timely, credible and useful for clinical-development decisions. Our work addresses challenges in precision oncology, pediatric cancers and rare diseases, where patient populations are often small, outcomes may mature slowly and evidence can be limited or heterogeneous.

Our research connects three complementary areas: efficient adaptive trial design and dose optimization; critical assessment and integration of external or counterfactual evidence; and dynamic monitoring that links the maturity of accumulating information to interim decisions. We develop open-source R packages and other practical tools to support transparent implementation and evaluation of these methods.

Close collaboration with St. Jude clinical investigators is central to our approach. Real trial questions motivate the methodology, and validated methods are translated back into study designs, protocols, statistical analysis plans, simulation studies and monitoring rules.

The Team

The Pan Group's team physicians and researchers by developing novel design methods for clinical trials and research initiatives.

Our research summary

Our central goal is to connect clinical-development questions with statistical designs, credible evidence and appropriately supported decisions. We study Bayesian and adaptive methods for dose finding and optimization, single-arm and randomized trials, time-to-event endpoints, delayed treatment effects and multi-arm or platform designs. We also investigate the qualifications and assumptions required to use external information in clinical research.

Across these areas, we emphasize operating-characteristic calibration, estimand alignment, reproducibility and the transparent communication of uncertainty. Our methods are motivated by collaborative research in pediatric solid tumors and other rare conditions, with a focus on practical implementation in clinical trials.

Representative research and software are organized below by scientific objective. These three areas are connected by a common focus on generating evidence that is fit for a particular clinical decision.

Efficient trial design and dose optimization

We develop efficient trial designs for small and heterogeneous patient populations. Topics include Bayesian and model-assisted dose finding and optimization, seamless early-phase development, single-arm and randomized phase II studies, biomarker-enriched studies and multi-arm platform trials. Our work on the Keyboard design includes model-assisted dose finding for drug combinations.

Our objective is to align trial design and operating characteristics with the decisions investigators must make, including dose selection, futility, advancement of candidate treatments and the addition of new study arms. 

PlatformDesign illustrates how preplanned experimental arms can be added during an ongoing multi-arm platform trial.

Representative open-source R packages

  • PlatformDesign
    Implements optimal two-period multi-arm platform designs that can add preplanned experimental arms while controlling specified error rates. 
  • UnifiedDoseFinding
    Implements and simulates dose-finding designs for nonbinary toxicity outcomes, including ordinal grades, weighted toxicity scores and continuous endpoints.  
  • MinEDfind
    Bayesian design for finding a minimum effective dose, with an interactive Shiny application.
  • Keyboard
    Model-assisted Bayesian dose-finding methods for single-agent and combination trials, including optimal biological dose designs.  
  • SAME
    Seamless adaptive multi-arm, biomarker-enriched phase II/III trial design with survival endpoints and sample-size re-estimation.  
  • SEARS
    Seamless dose escalation and expansion that combines toxicity-guided escalation with efficacy-based dose comparison. 
  • frequentistSSD
    Designs screened-selection phase II trials with survival endpoints, including arm selection, sample-size planning and empirical operating characteristics. 
  • frequentistSSDBinary
    Designs screened-selection phase II trials with binary outcomes, including two- or three-arm selection and empirical operating characteristics. 
  • BAR
    Supports outcome-adaptive Bayesian randomization and related trial-planning calculations. 

External and counterfactual evidence

Historical controls and other external data can be valuable when a conventional concurrent control group is difficult to obtain, but similarity of observed outcomes alone is not enough to justify borrowing. Our research addresses how the target clinical question, populations, endpoints, time zero and data quality should be assessed before external evidence influences a trial.

This program builds on work in calibrated Bayesian borrowing and now includes publicly available research on Qualify-Then-Borrow (QTB). We are investigating external controls, causal transportability, counterfactual benchmarks, estimand alignment and sensitivity analyses; these remain methodological research areas rather than claims of a fully validated or deployed clinical prediction system.

The central question is not how much external information can be incorporated, but whether it is sufficiently qualified and relevant to support the particular inference or decision.

Dynamic monitoring and adaptive decision-making

Evidence accumulates on different time scales in an ongoing trial. In pediatric oncology, endpoints such as progression-free and overall survival may require substantial follow-up, while decisions about futility, expansion or continued enrollment often cannot wait. We develop methods that make this tension explicit in prospective trial monitoring.

Our work includes Bayesian Go/No-Go rules, early futility monitoring, information-time alignment for time-to-event estimands, delayed treatment effects and non-proportional hazards, interim operating-characteristic calibration and adaptive decisions when realized enrollment differs from the original plan. These methods are designed to connect the maturity and credibility of the observed evidence with the decision being considered. 

Example software - DTE-BOP2: Bayesian Optimal Phase II design under Delayed Treatment Effects

DTEBOP2 illustrates simulation-based trial planning when treatment effects may be delayed.

Representative open-source R packages

  • BayesDesign
    Event-driven Bayesian two-stage single-arm phase II designs using time-to-event endpoints.  
  • BayesOrdDesign
    Bayesian group-sequential designs for ordinal endpoints, supporting proportional-odds, non-proportional-odds and model-switching approaches. Archived from current CRAN on 10 July 2025 (dependency archived).  
  • OneArm2stage
    Two-stage single-arm designs for time-to-event outcomes, including optimal, minmax and admissible options. 
  • UnplanSimon
    Adapts Simon's two-stage design to under- or over-enrollment through threshold and sample-size adjustments, with post-trial inference tools.  
  • DTEBOP2
    Bayesian randomized phase II trial design for delayed treatment effects, accounting for uncertainty about the delay in simulations and interim planning.  
  • WCRBayesDesign
    Calibrates window-cohort and controlled-roll-on Bayesian two-stage designs for single-arm trials with time-to-event endpoints.  
  • AHTauDesign R package (v1.0) 
    Supports prospective planning of two-arm time-to-event trials using average hazard estimands, evaluating clinically meaningful horizons, information support, sample size and follow-up; information support is distinct from statistical power.

From methods to evidence-informed clinical decisions

Our research is grounded in collaboration with clinicians and multidisciplinary trial teams. Statistical methods are translated into study designs, protocols, operating-characteristic simulations, interim monitoring procedures, statistical analysis plans and reproducible software. The aim is to provide both an appropriate method and a transparent basis for the clinical-development decision it supports.

A related direction is Evidence-to-Decision (E2D), which organizes statistical evidence around specific future development actions rather than a single generic declaration that a treatment is promising. We are also exploring carefully validated, AI-assisted workflows for organizing and reproducing evidence-generation tasks; scientific validity and consequential decisions remain subject to statistical evaluation, clinical judgment and human oversight.

Contact us

Haitao Pan, PhD
Associate Member, St. Jude Faculty
Department of Biostatistics
MS 768, Room R6030

St. Jude Children's Research Hospital

262 Danny Thomas Place
Memphis, TN, 38105-3678 USA
haitao.pan@stjude.org
262 Danny Thomas Place
Memphis, TN, 38105-3678 USA
GET DIRECTIONS