麻豆AV

Title:聽Multi-State Survival Models and Dementia

Abstract:聽Dr. Therneau grew up on a dairy farm in southeast Minnesota. He received his PhD from Stanford in 1983, and has been a member of the Biostatistics division at Mayo Clinic since 1985. He is coauthor of the book "Modeling Survival Data", a fellow of the American Statistical Association, and recipient of the 2020 W.J. Dixon award for excellence in statistical consulting. He first started writing S software to support his analyses in 1984, which effort grew into the survival package in R. He also enjoys woodworking, bicycling, and canoeing in the BWCA. Five grandchildren may be his greatest delight and reward.

I am a statistician who works in medical research, and I am also the author of the R survival package. This talk is a direct result of that three way interaction and also will have three intersecting threads. Medical question and data. What are the patterns of death and dementia in a general population, as a function of age and sex, and what is the impact of select other factors, e.g., amyloid accumulation, on these patterns? How do the factors interact in their effect? Data is from the Mayo Clinic Study of Aging, an age and sex stratified random sample from Olmsted County, Minnesota. The results tell a fascinating story. Statistical models. New problems drive new approaches. This analysis is based on multi-state hazard models, a fairly straightforward extension of the ordinary Cox model. In this analysis there are only 3 states (not demented, dementia, death) and three possible transitions between states. Nevertheless, this simple extension a provides much richer insight into the process. (In the last few years multi-state models have supplanted Cox models as my primary analysis tool.) Software. New methods create new software. The survival package is my "daily driver", so any new statistical method inevitably leads to an attempt to make the approach easy to use and reliable. In this case the target has been to make multi-state models as simple as coxph, and multi-state outcomes as easy to display as survfit. Software will only be touched on in this talk, however, with said material deferred to a later tutorial.