COMPUTATIONAL SOCIAL SCIENCE RESEARCH COLLOQUIUM /COLLOQUIUM IN COMPUTATIONAL AND DATA SCIENCES – Lifetime/Survival/Reliability/Duration Analysis for Computational Models – Robert Axtell
Computational Social Science Research Colloquium /
Colloquium in Computational and Data Sciences
Professor, Computational Social Science Program
Department of Computational and Data Sciences, College of Science
Department of Economic, College of Humanities and Social Sciences
George Mason University
Lifetime/Survival/Reliability/Duration Analysis for Computational Model
Friday, February 15, 3:00 p.m.
Center for Social Complexity Suite, 3rd Floor Research Hall
All are welcome to attend.
In a variety of computational models, structures arise, evolve, then disappear, perhaps replaced by other, comparable structures. For example, in some economic models firms form from the interactions of agents, operate for some period of time, and then exit. In housing models, households hold mortgages for finite periods of time before replacing them either due to refinancing or moving to a new house. In political (marketing) models the interests of parties (businesses) are aligned with certain segments of voters (consumers) for a period of time, until competition leads to realignment (brand switching). In environmental policy models specific polluting technologies have finite lifetimes and are eventually replaced by cleaner technologies. In disease models people are infected for varying lengths of time based on their health status, policies, etc. Traffic jams and conflicts have finite duration.
In this talk I will review the mathematical and statistical formalisms of lifetime analysis, also known as survival analysis in biostatistics and reliability analysis in engineering, focusing on the concepts most useful for computational models. Specifically while the former field has concerned itself with censored data (e.g., short clinical trials during which not all patient health outcomes can be observed), and the latter has focused on schemes to manage unreliable equipment, in computational modeling we often need to better understand both age and lifetime distributions of objects in our models, typically have large amounts of quasi-exhaustive ‘data,’ normally know some covariates, and usually work in discrete time.
I will go through the inter-relationships between survival, failure rate, and life expectancy functions, using parametric distributions to illustrate the main ideas. Then I will work through an extended example based on data concerning U.S. business firms, focusing on the connections between firm age and lifetime distributions, which ends with somewhat surprising conclusions, due to high failure rates among young firms (high ‘infant mortality’).