Xiaoge Zhang, PhD
HEOR Technical Portfolio & Knowledge Base
Bridging Academic Econometrics with Industrial HTA Standards
A professional showcase of high-fidelity health economic modeling, Bayesian evidence synthesis, and real-world causal inference, tailored for advanced HTA applications and NICE/SMC submissions.
π Flagship Case Studies
Replicating a Published Multinomial NMA
Response rates in relapsed/refractory myeloma β van Beurden-Tan et al. 2022 (BMC Cancer, CC-BY)
A Bayesian multinomial network meta-analysis over 17 phase III trials and 16 treatments, rebuilt in Stan from the published trial counts and cross-checked by running the original WinBUGS code in JAGS. All sixteen response rates reconcile to within 0.27 percentage points and SUCRA to within 0.2 β but only after correcting the data file the paper supplies for replication, which produces an error of 19.6 points as published. The rebuild then asks what the ranking rests on: the treatment reported as best holds 47% of the posterior probability of being best, and loses that place when an undiscussed continuity constant or the prior width is changed.
NICE TA1092 (Pembro) Case Study
Mock Case β Extrapolation Audit & Spline Sensitivity
An EAG-style forensic audit of survival extrapolations (standard parametric vs. flexible splines), treatment-effect waning (5-7 years), and curve-crossing capping, using Guyot-reconstructed patient-level data.
NICE TA975 (CAR-T) Case Study
Full Submission Replication
A complete end-to-end reconstruction of a CAR-T therapy appraisal (Tisagenlecleucel). Features state-transition modeling, parametric survival fitting, and ICER calculations via Patient Access Schemes (PAS).
ML-NMR: AML Population Alignment (TA1013)
Mock Case β Population Alignment & Synthesis
A synthetic-data deep dive into Multilevel Network Meta-Regression. Solves the βExtrapolation Gapβ using 64-point Sobol integration and Cholesky decomposition to bridge IPD and AgD seamlessly in a Bayesian framework.
RWE External Control Arm: R/R DLBCL
Mock Case β Doubly Robust Causal Inference Pipeline
A synthetic-data methodological showcase of External Control Arm construction for a fictional NICE appraisal. Implements IPTW, TMLE (doubly robust estimation), and E-value sensitivity analysis β aligned with the NICE RWE Framework (2022).
PSM: Survival Extrapolation in RRMM
Mock Case β Treatment Effect Waning & Recursive Hazard Chaining
A synthetic-data deep dive into long-term survival extrapolation in a Partitioned Survival Model. Contrasts Naive HR Switching (the Hazard Cliff) against Recursive Hazard Chaining, with full cost-effectiveness analysis and PSA.
π Replication & Verification
Published models rebuilt from their own papers, reproduced number by number, and then checked β against the source data, against the reporting, and against the verification standards the field already has.
Verifying a Published Hypertension Markov Model
Cost-effectiveness of two antihypertensive regimens in China β Feng et al. 2023 (J Int Med Res, CC BY-NC)
A four-state Markov cost-effectiveness model rebuilt in R from the published inputs and checked stage by stage against the TECH-VER verification checklist. The published QALYs reproduce exactly and both armsβ costs reconcile β the amlodipine arm only once its drug price is taken from the paperβs own Figure 4 rather than its Table 2, which differ by 35 CNY a year. The trial percentages behind the transition matrix are cumulative over a 36-month follow-up and are used as one-year probabilities: run the model over the trialβs own three years and it returns 11.2% cumulative death where 4.1% went in. Correcting that halves the health gain, and the direction of the published conclusion survives while its magnitudes do not.
Rebuilding a Published Budget Impact Model
Venetoclax in AML β a replication of Palacios et al. 2024 (PLoS One, CC-BY)
A three-year budget impact model for venetoclax combinations in acute myeloid leukaemia, reconstructed from Palacios et al. (PLoS One 2024, CC-BY) rather than restated: epidemiology funnel, market share in both worlds, six cost components built from published quantities, and a one-way sensitivity analysis run on the parameters the source itself varied. Every input names where it came from β a trial, the Delphi panel, a national price list, or an explicit assumption β and the year-3 result reconciles to within 0.1% of the published figure under both payer perspectives.
A working deliverable would be an Excel workbook; this page exists to make the structure and its judgements visible.
π The Technical Library
NMA & Evidence Synthesis Library
A comprehensive suite of technical guidelines and code templates for diverse Network Meta-Analysis challenges.
Population Adjust
Advanced Synthesis
Real-World Evidence & Causal Inference Library
Methodological guidelines and implementation templates for causal inference, confounding adjustment, and external control arms.
RWE Methodology
A focused library of methodological frameworks, implementation guides, and causal inference templates for observational studies, aligned with the 2022 NICE RWE Framework.
π οΈ Tech Stack & Methodologies
R (HEEMOD, multinma, tidyverse) Stan / Bayesian Inference Excel (VBA / PowerQuery) NICE DSU TSD Guidelines SuperLearner / TMLE