EXTENSION OF AGENT BASED MODELING IN CHINA’S NATIONAL VOLUME BASED DRUG PROCUREMENT: BIDDING SIMULATION AND POLICY IMPLICATIONS
Author(s)
Qian Xu, Master1, Jia WANG, Master2, Zhuangqi Li, Master1, Xuanwen Ding, Master3, chao he, PhD4, Zhongyu Wei, PhD5, Bao Liu, PhD1;
1School of Public Health, Fudan University, Shanghai, China, 2Shanghai Innovation Institute, Shanghai, China, 3School of Data Science, Fudan University, Shanghai, China, 4Shanghai Key Laboratory of Rail Infrastructure Durability and System Safety, Tongji University, Shanghai, China, 5School of Data Science , Fudan University, Shanghai, China
1School of Public Health, Fudan University, Shanghai, China, 2Shanghai Innovation Institute, Shanghai, China, 3School of Data Science, Fudan University, Shanghai, China, 4Shanghai Key Laboratory of Rail Infrastructure Durability and System Safety, Tongji University, Shanghai, China, 5School of Data Science , Fudan University, Shanghai, China
OBJECTIVES: To extend agent based simulation model for competitive bidding in China's National Volume Based drug Procurement (NVBP) and explore its policy implications for procurement optimization.
METHODS: We developed a multi agent simulation framework calibrated with real-world data from NVBP rounds 2-9 (excluding Round 6), involving 328 drugs and 2,226 enterprises. A rule based Agent Based Model (ABM) was constructed to simulate bidding of firms under rules. To address the limitation that rule based ABMs rely on random single step actions, we integrated a Markov Decision Process (MDP). Reinforcement Learning (RL), Large Language Models (LLMs), and markup-based pricing algorithms were employed to enable firms to iteratively learn profit maximizing strategies. Sensitivity analyses varied key policy parameters (maximum valid bid price, agreed procurement ratio) and market parameters (agreed procurement volume, actual procurement volume, unit production costs).
RESULTS: The rule based ABM demonstrated good fidelity in reproducing real world outcomes (overall RMSE 12.86%). The simulation accuracy varied across different rounds in which Round 7 showed relatively poorer fit (RMSE 15.27%) likely due to unique market conditions, while Round 9 achieved better performance (RMSE 9.84%). In the MDP setting, firms successfully acquired optimal strategies through iterative learning. Simulated bids strongly correlated with actual bids (Spearman ρ=0.85-0.88; p<0.001), with RL consistently outperforming LLM and markup approaches regarding firm profitability. Sensitivity analyses revealed the maximum valid bid price as the most influential policy parameter affecting bidding behavior, whereas the agreed procurement ratio had limited impact. Profits were highly sensitive to deviations between actual and agreed procurement volumes.
CONCLUSIONS: While the agent based modeling approach effectively reproduced real-world bidding outcomes, the MDP extension enhanced simulation realism and presented interpretable evidence of firms’ strategic adaptations. These findings highlighted that accurate demand forecasting and rational rule design of maximum bid price were key policy levers for optimizing NVBP in China.
METHODS: We developed a multi agent simulation framework calibrated with real-world data from NVBP rounds 2-9 (excluding Round 6), involving 328 drugs and 2,226 enterprises. A rule based Agent Based Model (ABM) was constructed to simulate bidding of firms under rules. To address the limitation that rule based ABMs rely on random single step actions, we integrated a Markov Decision Process (MDP). Reinforcement Learning (RL), Large Language Models (LLMs), and markup-based pricing algorithms were employed to enable firms to iteratively learn profit maximizing strategies. Sensitivity analyses varied key policy parameters (maximum valid bid price, agreed procurement ratio) and market parameters (agreed procurement volume, actual procurement volume, unit production costs).
RESULTS: The rule based ABM demonstrated good fidelity in reproducing real world outcomes (overall RMSE 12.86%). The simulation accuracy varied across different rounds in which Round 7 showed relatively poorer fit (RMSE 15.27%) likely due to unique market conditions, while Round 9 achieved better performance (RMSE 9.84%). In the MDP setting, firms successfully acquired optimal strategies through iterative learning. Simulated bids strongly correlated with actual bids (Spearman ρ=0.85-0.88; p<0.001), with RL consistently outperforming LLM and markup approaches regarding firm profitability. Sensitivity analyses revealed the maximum valid bid price as the most influential policy parameter affecting bidding behavior, whereas the agreed procurement ratio had limited impact. Profits were highly sensitive to deviations between actual and agreed procurement volumes.
CONCLUSIONS: While the agent based modeling approach effectively reproduced real-world bidding outcomes, the MDP extension enhanced simulation realism and presented interpretable evidence of firms’ strategic adaptations. These findings highlighted that accurate demand forecasting and rational rule design of maximum bid price were key policy levers for optimizing NVBP in China.
Conference/Value in Health Info
2026-05, ISPOR 2026, Philadelphia, PA, USA
Value in Health, Volume 29, Issue S6
Code
HPR5
Topic
Health Policy & Regulatory
Topic Subcategory
Pricing Policy & Schemes, Procurement Systems
Disease
STA: Generics