报告题目:Model-Based Data Analysis: Detecting and Resolving Nonidentifiability
报告人:Michael Li,University of Alberta
报告时间:2016年12月30日16:00
报告地点:复杂系统研究所报告厅
报告摘要:
When a mathematical model is confronted with data, many sensitive parameters can not be estimated directly from the data and need to be estimated indirectly through model fitting. One of the main challenges in model fitting is the nonidentifiability issue: infinitely many parameter values can produce the same quality fit. This may not sound like a serious issue since we can use any of these values. However, if the goal of modeling is to estimate quantities that are not observable, then it is very often the case that two different parameter values with the same model fit to data can produce drastically different estimations on unobservable quantities. It is essential to detect and resolve nonidentifiability when performing model fitting. After introducing different notions of nonidentifiability, I will review some existing methods for detecting and ranking nonidentifiable parameters. I will introduce a new method using singular value decomposition and variance decomposition, which has several advantages over existing methods. I will then use model-based HIV estimation as an example to illustrate the issues and their solutions.