Florence Nightingale Colloquium presents Hein Putter
- Friday 24 September 2021
- The seminar is targeted at a broad audience, in particular we invite master students, PhD candidates and supervisors interested or involved in the Data Science Research programme as well as colleagues from LIACS and MI to attend. The seminar is organized by the DSO, MI and LIACS.
- Kaltura Live Room
Immortal time bias, dynamic prediction and landmarking
This talk will give a general introduction to two topics, immortal time bias and dynamic prediction, and one concept linking the two, landmarking. I will use examples from the medical field for illustration.
Topic 1 is immortal time bias.
Did you know that people who eat more birthday cakes and drink more birthday beers live longer? This is not because eating cakes and drinking beers is particularly healthy, but it is because people living longer will have had more opportunity for eating birthday cakes and drinking birthday beers. It is an example of so-called immortal time bias, and there are many (more serious) examples of it in the medical literature.
Topic 2: Dynamic prediction. Prediction models play a very prominent role in medical research. The vast majority are based on a fixed starting time, for instance diagnosis of a disease or start of treatment. But in many situations obtaining a prognosis for the patient is also important at later time points, when more information on the clinical status of the patient has become available, in the form of so-called time-dependent covariates. The prediction models should be updated based on that information, leading to dynamic prediction models.
The concept linking the two is landmarking. The landmark concept was introduced in 1983 by Anderson et al. as a way to correct for the immortal time bias when assessing the effect of response to treatment on survival. The landmark approach was revived by van Houwelingen in 2007 to obtain dynamic predictions when time-dependent covariates are present. Landmarking avoids specification of complex joint models for time-dependent covariates and survival and yields predictions that are robust against violations of model assumptions like the proportional hazards assumption. This talk will give a general introduction and illustrates the use of landmarking to obtain a dynamic prediction model for breast cancer patients.
Join the webinar via Kaltura Live Room
Kaltura Live Room works best in Edge, Chorme and Firefox. Make sure you activate your camera and microphone beforehand in order to interact with the speaker and participate in discussion. The room opens for the public at 12:50.Register for the Kaltura Live Room link