BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Causal discovery with tiered background knowledge - Christine Bang
  (Københavns Universitet (University of Copenhagen))
DTSTART:20260304T103000Z
DTEND:20260304T111500Z
UID:TALK244390@talks.cam.ac.uk
DESCRIPTION:Causal graphs are useful tools for understanding complex causa
 l structures and can help us deciding on interventions or identifying caus
 al effects. Causal graphs are usually constructed based on expert knowledg
 e\, which may be insufficient or incorrect. Causal discovery methods const
 ruct causal graphs in a data-driven way and serve as an alternative to the
  classical expert-driven approach but come with other limitations. Typical
 ly\, without strong assumptions\, we cannot identify all causal directions
 . Moreover\, causal discovery methods are sensitive to errors based on e.g
 . statistical testing when applied to real data.&nbsp\;In practice\, we of
 ten have more causal information available than what can be obtained from 
 the data alone\, and the use of so-called background knowledge in causal d
 iscovery is a way to bridge the purely data-driven and the purely expert-d
 riven approaches. Time structure induces a partial causal ordering of the 
 variables\, which I will refer to as tiered background knowledge. This typ
 e of background knowledge is common\, e.g. for cohort data\, and it improv
 es causal discovery methods such that the output becomes more informative 
 and reliable.In this talk\, I will first explain how tiered background kno
 wledge can be incorporated in constraint-based causal discovery algorithms
 . Then\, I will show how this improves graphs estimated using finite sampl
 e data. In addition\, I will show which time structures yield the most inf
 ormative graphs and that the output graphs have some desirable theoretical
  properties. Finally\, I will show in which ways tiered background knowled
 ge is useful if we allow for unobserved confounding and consider multiple 
 datasets with overlapping variables.&nbsp\;
LOCATION:Seminar Room 1\, Newton Institute
END:VEVENT
END:VCALENDAR