Program
Invited Speaker
Morten Overgaard (Aarhus University, Denmark)
Regression analysis with jack-knife pseudo-observations
More than twenty years ago, a pseudo-observation approach to regression in survival analysis was suggested. Such an approach allows for handling incomplete observation of a relevant outcome variable by transforming the available data into pseudo-observations that can replace the potentially unobserved outcomes in a regression analysis. The pseudo-observations are here the jack-knife leave-one-out pseudo-values from a suitable estimator.
The original motivation of the approach was in a multi-state setting with right censoring. It has since been suggested for settings with complications such as left-truncation, interval censoring or recurrent events.
We will review this sort of approach and attempt to find answers to questions such as: How do I use it? When does it work? Why is left-truncated data a challenge? Is variance estimation an issue? How does it compare to similar approaches? Where is it going from here? Are pseudo-observations fun?
Nan van Geloven (Leiden University, The Netherlands)
Causal prediction of time-to-event outcomes
A key aim of clinical prediction models (or prognostic models) is to provide individualized risk estimates that support patients and doctors in making treatment decisions. However, as most prediction models are derived from observational data where some individuals already received the treatment the model aims to inform, standard prediction methods often fail to provide the necessary information for this.
Causal predictions (also called counterfactual predictions or predictions under intervention) have recently gain traction as an alternative way to support treatment decisions. These are estimates of risks under specified treatment strategies, for example a patient’s baseline risk `if they do not initiate the treatment’ or the risk ‘if they do initiate the treatment’. A major challenge in estimating and evaluating causal predictions in observational data is confounding adjustment.
In this talk, I will outline how causal inference methods such as marginal structural models and the clone-censor-reweight approach can be adapted for the purpose of developing and evaluating causal predictions of time-to-event outcomes.
Dennis Dobler (Dortmund University, Germany)
Resampling options in survival and event history analysis
Resampling plays a versatile role in every branch of statistics. For example, it is used to compute variance estimates, to avoid the suboptimal normal distribution approximation in inference methods, or as bagging (bootstrap aggregating) in random forest algorithms. There exists an abundance of resampling methods, in particular, the classical, weighted, multiplier, and parametric bootstraps, and random permutation.
In survival analysis, the issue of incomplete data demands special attention when it comes to resampling. For instance, Akritas (1986, JASA) found that Efron’s suggestion (1981, JASA) to draw with replacement from the censored data points before recomputing the Kaplan-Meier estimator allows the construction of asymptotically valid confidence bands for the survival function. And also that, in contrast, Reid’s approach (1981, Biometrika) to independently resample from the Kaplan-Meier curve alters the covariance structure of the resampled Kaplan-Meier estimator; hence, it should not be used for constructing such confidence bands.
This talk will provide an incomplete overview of various resampling procedures in different survival analytic applications, address some requirements and intuitions behind these procedures, and also discuss some computational aspects.
