Dina Research School
Workshop:
Tired of waiting? - an introduction to statistics for
time to event data
Koldkærgård Konferencecenter, 21-22 October 2004
Preliminary programme
Thursday, October 21
- 11:00
- Arrival and accommodation
- 12:00
- Lunch
- 13:00
- Introduction and presentation of participants
Anders Ringgaard Kristensen, The Dina Research School
- 13:15
- Theory session I: Introduction
Torben Martinussen, KVL.
Analysis of survival data or more generally waiting time data calls for
special statistical methods. This is because the exact waiting times are often
not fully observed. The most common case is right-censoring meaning that it is
only known that the waiting time exceeds an upper limit. The special character
of waiting time data is introduced and it is shown that traditional
statistical methods will fail. This is shown in relation to various datasets.
Finally the Kaplan-Meier estimator of the survival function is introduced.
- 14:00
- Short break followed by computer exercises.
- 15:00
- Coffee break
- 15:30
- Theory session II: Nonparametric methods.
Torben Martinussen, KVL.
Classical nonparametric methods to the analysis of censored observations will
be described. These include the Kaplan-Meier estimator that estimates the
survivor function and the log-rank test that may used to compare the waiting
time distributions of two (or more) groups. The methods are nonparametric and
as such appealing.
- 16:15
- Computer exercises
- 17:00
- Theory session III: Cox-regression part I.
Torben Martinussen, KVL.
In practise one often has several explanatory variables that might influence
the response variable (waiting time) so a regression model is needed. The
absolute dominant model in this area is Cox's proportional hazards model. It
is very flexible as only a part of the waiting time distribution is specified
conditional on the explanatory variables. The model is introduced and its
special structure is explained. It is shown how to fit the model in R.
- 18:00
- Dinner
- 19:00
- Case study I: Survival analysis in animal breeding
Lars Damgaard, Animal breeding and Genetics, Research Centre Foulum
Survival models were introduced in the area of animal breeding in 1984 to
study longevity of dairy cows. The trait considered was time from first
calving until culling. Today this trait is included in several breeding
programs for dairy cows. Recently, survival models have also been used to
study additive genetic aspects of resistance to diseases in growing pigs, beef
bulls and fish.
In this presentation, I will first shortly present the survival traits
considered in animal breeding. Secondly, I will describe the underlying
genetic model (the additive genetic infinitesimal model), and the associated
proportional hazards model applied in animal breeding. Finally, I will present
results from a joint genetic analysis of calving difficulty and longevity of
dairy cows. In this study, calving difficulty was recorded as a categorical
trait taking values in one out of 3 ordered categories. Longevity was recorded
as a survival trait defined as time from first calving until culling. Focus
will be on results obtained for longevity.
- 19:45
- Computer exercises
- 21:45
- Coffee and sandwich
Friday, October 22
- 7:45
- Breakfast
- 8:30
- Discussion of computer exercises
- 9:00
- Case II: Applications of survival analysis in behavioral research
Karen Thodberg and Mette Herskin, Animal Health and Welfare,
Research Centre Foulum
In the presentation we will describe a number of research projects where
survival analysis has been used to analyse behaviour data. The application of
survival analysis methods is illustrated with two case stories concerning
nursing motivation in sows (Cox regression) and a laser test for measuring
pain sensitivity in cows. We conclude by discussing how we would like to
improve the analyses by taking into account repeated measurements and time
dependent covariates.
- 9:45
- Coffee break
- 10:00
- Theory IV: Cox-regression part II and more specialised topics.
Torben Martinussen, KVL.
Although the Cox-model is very flexible there are some assumptions underlying
the model. These assumptions are discussed and it is shown how to investigate
whether they are fulfilled in practise. Some explanatory variables are special
in that they may be thought of as block variables. In ordinary linear models
their effects are typically modelled as random effects. Due to the
non-linearity of the models used in this field I will argue that their impact
are often modelled more naturally in a different way using so-called marginal
models.
- 11:00
- Break
- 11:05
- Computer exercises
- 11:45
- Discussion and closing
Anders Ringgaard Kristensen, The Dina Research School
- 12:00
- Lunch and departure
Author: phd@dina.kvl.dk. Updated:
17 oktober 2004